Home Search Collections Journals About Contact us My IOPscience

Uncertainty on PIV mean and fluctuating velocity due to bias and random errors

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2013 Meas. Sci. Technol. 24 035302

(http://iopscience.iop.org/0957-0233/24/3/035302)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 157.89.65.129
This content was downloaded on 12/02/2015 at 11:49

Please note that terms and conditions apply.

IOP PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY
Meas. Sci. Technol. 24 (2013) 035302 (15pp) doi:10.1088/0957-0233/24/3/035302

Uncertainty on PIV mean and fluctuating

velocity due to bias and random errors
Brandon M Wilson 1 and Barton L Smith 2
1
Los Alamos National Laboratory, Los Alamos, NM 87544, USA
2
Mechanical and Aerospace Engineering, Utah State University, Logan, UT 84322, USA
E-mail: [email protected] and [email protected]

Received 28 August 2012, in final form 10 January 2013

Published 4 February 2013
Online at stacks.iop.org/MST/24/035302

Abstract
Particle image velocimetry is a powerful and flexible fluid velocity measurement tool. In spite
of its widespread use, the uncertainty of PIV measurements has not been sufficiently addressed
to date. The calculation and propagation of local, instantaneous uncertainties on PIV results
into the measured mean and Reynolds stresses are demonstrated for four PIV error sources that
impact uncertainty through the vector computation: particle image density, diameter,
displacement and velocity gradients. For the purpose of this demonstration, velocity data are
acquired in a rectangular jet. Hot-wire measurements are compared to PIV measurements with
velocity fields computed using two PIV algorithms. Local uncertainty on the velocity mean and
Reynolds stress for these algorithms are automatically estimated using a previously published
method. Previous work has shown that PIV measurements can become ‘noisy’ in regions of
high shear as well as regions of small displacement. This paper also demonstrates the impact
of these effects by comparing PIV data to data acquired using hot-wire anemometry, which
does not suffer from the same issues. It is confirmed that flow gradients, large particle images
and insufficient particle image displacements can result in elevated measurements of
turbulence levels. The uncertainty surface method accurately estimates the difference between
hot-wire and PIV measurements for most cases. The uncertainty based on each algorithm is
found to be unique, motivating the use of algorithm-specific uncertainty estimates.

Keywords: PIV, uncertainty

(Some figures may appear in colour only in the online journal)

1. Introduction paper demonstrates the method proposed by Timmins [1] and

Timmins et al [2] for velocity and Reynolds normal stress
Particle image velocimetry (PIV) offers significant advantages profiles based on real data. Results and discussion are limited
over traditional, point-wise velocity measurement techniques. to two-component single-camera digital-PIV, although much
PIV is non-intrusive, can be performed over a wide range of of the results are important to stereo and micro PIV.
spatial scales, and provides a spatial velocity field at an instant
in time. The technique works well in multiple flow regimes 1.1. PIV accuracy
and media. While originally thought of as a flow visualization There are dozens of error sources in a PIV measurement.
method, improvements in PIV algorithms have advanced the Errors can stem from the system calibration, timing errors,
method to an accurate measurement technique. Unfortunately, the optical setup, particle slip, as well as other systematic and
PIV accuracy varies in time and space, and in a manner that random sources. The effects of many of these are discussed
is not a function of the measured velocity. While there has in some detail in the recent book by Adrian and Westerweel
been considerable research on PIV error sources, methods of [3] based on their earlier work. As PIV algorithms become
quantifying the effects of many of these error sources through more complex, including methods that alter the measurement
uncertainty quantification have not existed until recently. This according to local flow conditions, the accuracy of PIV

0957-0233/13/035302+15$33.00 1 © 2013 IOP Publishing Ltd Printed in the UK & the USA
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

becomes more difficult to pin down, as discussed by Stanislas Rabinovich [17]. Uncertainty implies that one can make a
et al [4]. general statement (usually probability based) about the range
Many previous papers address the accuracy of PIV within which one expects to find the true error, while studies on
algorithms, examining the bias and/or random (often called accuracy endeavor to demonstrate the size of the error under
rms) error. The following examples are a small sample. specific conditions.
The bias from loss-of-correlation that comes with particle Random error always leads to precision uncertainty on
displacement and the means to correct it was demonstrated time-averaged quantities for a finite number of samples.
by Keane and Adrian [5]. A broader study of the bias and Precision uncertainties present some unique issues in PIV
noise generated by PIV algorithms was presented in Huang measurements, since some data may be spatially independent
et al [6]. The ‘peak-locking’ effect, and how it stems from while others are not [18]. Many PIV studies cite precision
bias and random errors was studied by Gui and Wereley uncertainty on mean quantities, but precision uncertainty of
[7]. They also suggested ways of mitigating these effects. quantities such as Reynolds stresses are usually ignored.
Nogueira et al [8] discussed how peak-locking can lead to A rigorous Taylor series method (TSM) of the uncertainty
errors in turbulence level measurements and how random error of the time-averaged velocity as well as the Reynolds stresses
contribution from peak locking can be assessed by making was performed by Gui et al [19]. Estimates of random error in
repeated measurements of the same flow using various t this study were made based on the PIV algorithm performance
values. This strategy is employed in the present work. for simple displacement, which is similar, but more limited,
Several studies, including Westerweel [9], have shown to the current approach. A study aimed at the uncertainty of a
that flow shear broadens the correlation peak, and therefore measurement of entropy generation [20] used a fixed value for
increases the random error in the measurement. These effects displacement uncertainty provided by Dantec3 of 0.03 pixels.
can be mitigated by deforming the interrogation windows The combination of several uncertainty sources was
(Scarano [10]). However, deforming the interrogation window considered by Lazar et al [21]. The analysis included such
requires resampling of the original image, which certainly sources as the laser timing error and particle slip as well
generates other errors. as the PIV algorithm. These authors considered only errors
The impact of some of these error sources can be on the time-averaged flow, and took the interesting approach
exacerbated by poor image quality. Many studies have reported of generating synthetic images based on the measured flow,
ways to improve PIV images with the aim of improving and then computed the difference between the original PIV
accuracy including Roth and Katz [11]. The presence of solid data and the vectors generated with the synthetic data as their
objects in the flow field generates errors for interrogation measure of uncertainty.
regions which straddle the object. A means to move from error to uncertainty is suggested by
Nobach and Bodenschatz [12] showed that the limit many studies on the effects of various PIV error sources, which
of PIV accuracy is much larger than often claimed (near are typically based on synthetic images or theory. Adrian and
0.1 pixels) due to illumination variations that occur due Westerweel [3] discuss common error sources and provide
to the laser sheet shape coupled with out-of-plane motion. analytical descriptions of how these affect the PIV calculation
These effects were particularly pronounced in cases where the (e.g. correlation peak). The effects of the same error sources
particle images overlapped. of interest in the present work were studied for an undisclosed
The accuracy for PIV, single-pixel PIV and particle algorithm in Raffel et al [22].
tracking velocimetry (PTV) near a wall was recently In order to make use of these results, it is necessary
investigated by Kaehler et al [13]. These authors were able to (1) be able to locally quantify the error contributors and
to show that single-pixel PIV provides a large improvement (2) have knowledge of how a particular combination of error
over standard PIV, although, in most cases, PTV was superior. contributor values influence the bias and random uncertainty.
The impact of random error is mitigated by improving It is important to note that the effects of individual error
the dynamic velocity range (DVR) of the measurement, which sources on PIV results are strongly dependent on the PIV
is defined by Adrian [14] as the maximum measurable velocity algorithm. This is demonstrated in the results of the first three
divided by the minimum resolvable velocity. Poor dynamic PIV challenges [4, 23, 24], particularly figures 37 and 38 in
range is one of the primary limitations of PIV, but there have the third PIV challenge [4]. Thus, accurate PIV uncertainty
been several advances in PIV algorithms that significantly estimates may require algorithm-specific methods.
improve dynamic range. The primary advance in PIV dynamic PIV errors, and thus uncertainties, vary within a
range is window shifting [15]. More recently, advances in the measurement field due to local changes in the error
speed of PIV recording have made it possible to generate a contributors. Consequently, universal uncertainty estimates
dataset that is a composite of vectors computed from images (regardless of algorithm) are inappropriate. Two efforts have
acquired at different values of t, such that slower regions use been made recently to calculate uncertainties locally.
larger values of t [16]. This procedure makes determination The more recent effort is based on the hypothesis that the
of temporal resolution much more complex. correlation peak contains information about the uncertainty.
Charonko et al [25] showed that for the robust phase
1.2. PIV uncertainty correlation [26], the ratio of the largest correlation peak to
While the topics of accuracy and uncertainty are closely 3 Dantec Dynamics A/S, Tonsbakken 16-18, PO Box 121, DK-2740

related, the terms are not synonymous as pointed out by Skovlunde, Denmark.

2
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

the next largest is well correlated to the uncertainty in the Reynolds stresses; however, this is beyond the scope of this
measurement. This method has low cost and does not require study.
or provide information about the source of the uncertainty. Noting that any component of random measurement error
Timmins [1] and Timmins et al [2] have studied the effects contributes to the size of ŭ (and, similarly, to the size of the
of specific error sources (particle image density, diameter, total fluctuations), assessment of the random errors can be used
and displacement, and flow shear) on specific PIV algorithm to estimate the bias in the Reynolds stress. Often, the total
uncertainty using synthetic PIV images. The individual and measured fluctuations of a flow are reported as turbulence
correlated effects of these sources were determined to find the statistics without regard to the uncertainty generated from
PIV uncertainty under local conditions. By quantifying error random measurement errors. When the effect of random
source magnitudes, uncertainties were found for each vector measurement errors are not considered, an elevated level of
(in space and time) in a measurement set. They reported that turbulence is reported in all but the most ideal situations.
these four parameters could each generate bias and random In the work of Timmins et al [2], a method for calculating
error in the measurement, similar to the results of Raffel et al the instantaneous systematic and random measurement
[22] and others. While four uncertainty sources were studied, uncertainties for PIV algorithms was developed. This method
the method, in principle, allows for any uncertainty source has been implemented for the PRANA algorithm from Virginia
that can be measured. This same method is used in the present Tech [26, 28] and LaVision’s DaVis 7.2 [29]. In a subsequent
work and is described in more detail below. work, Wilson and Smith [27] propagated the instantaneous
Most flows of interest will generate varying levels systematic and random uncertainties into measured mean
of uncertainty throughout the measurement field for PIV and variances using both the Monte Carlo and Taylor-series
measurements. For example, consider a planar jet formed by methods.
a laminar channel flow. The flow is characterized by regions
of high shear at the edges of the jet and a core flow with a 1.4. The present study
nearly uniform velocity profile. The high shear edge regions
generate large amounts of ‘noise’ (random error) in the PIV In the present work, the methods introduced by Timmins
measurement [2, 22], while smaller, systematic uncertainties et al [2] and Wilson and Smith [27] are used to estimate the
are found within the core. Additionally, fluctuations in the uncertainty of PIV results affected by specific error sources
shear layer introduce temporal variation in the magnitude using real data. The experimental cases studied are designed to
of the shear in the shear layer. Generally, error sources (i.e. not only demonstrate components contributing to PIV errors,
particle image density, diameter and displacement) may also but to also demonstrate and quantify the random measurement
vary locally both temporally and spatially. errors and their effects on velocity means and fluctuations
on real PIV image sets. The PIV error sources considered in
this paper are the same as those demonstrated using synthetic
1.3. Uncertainty of Reynolds stresses
images by Timmins et al [2]: particle image density, particle
While uncertainties on the mean due to velocity fluctuations image diameter, displacement and gradients within the flow.
and PIV noise may be reduced by acquiring more data, This paper will focus on errors leading to uncertainties that
measurements of Reynolds normal stresses (which are the stem from the vector computation once images are acquired.
variance of the velocity) will always be biased upward due In what follows, ‘error sources’ will be limited to sources
to random measurement error. It may be assumed that at stemming from the vector computation. The uncertainty
any instant, the measured flow quantity is composed of four estimates will be based on the uncertainty surface method
components: the mean u, the actual velocity fluctuation u , [2]. A brief synopsis of the method will now be provided.
systematic error βu and random error ŭ. The measured or total For a specific PIV algorithm (meaning software and
fluctuations may then be written as û = u + ŭ. Using this its settings), a series of tests is used to determine the
relation, the measured velocity (u) at any time and location uncertainty response of the algorithm to particle displacement,
is obtained through the sum of the mean and measured particle image diameter, particle image density and shear.
fluctuations components, The tests consist of computing vector fields from a large
series of synthetic images that are based on flows with
u = u + û = u + u + ŭ + βu . (1) varying displacement and shear. The images also represent
In the remainder of the paper, Reynolds-averaged terms will a range of particle image density and particle image diameter.
be denoted as u, ûû, u u and ŭŭ. The quantity u u will be By comparing the resultant vector fields to the known
referred to as the Reynolds normal stress. displacement for each case, the distribution of error, or the
Wilson [27] discussed the effects of the instantaneous uncertainty for a specific combination of the four parameters
systematic and random measurement error on the measured can be determined. The four-dimensional uncertainty response
mean and variance through Reynolds-averaging. The variance is termed the ‘uncertainty surface.’ This process must be
and co-variances are synonymous with the Reynolds stresses. performed for each algorithm of interest, but the resultant
Random fluctuations in the instantaneous systematic and uncertainty surface is used for all future calculations.
random PIV measurement errors elevate the measured time- As part of the present work, a four-dimensional
averaged Reynolds normal stresses. Spatially unresolved uncertainty surface has also been generated for the DaVis
turbulent fluctuations can also cause an underestimate of 7.2 algorithm with the same processing parameters used

3
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

Figure 1. The facility consists of four components: (1) blower, (2) diffuser, (3) flow conditioning and the (4) contraction. Hot-wire and PIV
measurements were acquired downstream of the jet exit. A stepping motor traverse positioned the hot-wire probe. PIV measurements were
obtained in the PIV measurement region.

by Timmins et al [2] and using the same synthetic Wilson and Smith [27] account for this in the Reynolds normal
images. Uncertainties on the measured mean and Reynolds stress uncertainty formulation
normal stress are calculated by propagating the instantaneous

uncertainties using the equations derived by Wilson and Us+2 = p2û2 =

pû2
(4)
û
Smith [27]. and

Once a PIV dataset has been processed to find the velocity  2
 1 
N
Us2 = 
field, the values of the four error sources are determined −
U2 + p2û2 , (5)
from the raw image data and the velocity vectors. Then, û N − 1 i=1 ûi
the uncertainty surface is queried for the value of the
uncertainty for each instantaneous vector. Each component where Us+2 and Us−2 are the uncertainties on the Reynolds normal
û û
of velocity is assigned four instantaneous uncertainties by stress in the positive and negative directions, respectively, and
the uncertainty surface: a systematic uncertainty bu , an upper Uû2 is the total uncertainty on the instantaneous fluctuations
and a lower random uncertainty (ru+ and ru− ), and a random squared, which is calculated as

fluctuation uncertainty rû2 . The two-sided random uncertainty Uû2i = (bui − bu )4 + rû22 . (6)
allows different random uncertainties in the positive and i

negative directions due to asymmetric distributions. The The experimental cases studied are designed to not only
random fluctuation uncertainty is the uncertainty used in demonstrate components contributing to PIV errors, but to also
the calculation of the variance uncertainty [27]. As with demonstrate and quantify the random measurement errors and
the other uncertainty surface random uncertainty values, the their effects on velocity means and Reynolds normal stresses
random fluctuation uncertainty is obtained via instantaneous on real PIV image sets. In what follows, the experimental
PIV results from Monte Carlo simulated synthetic images. facility used in this study is described. This is followed by
The random fluctuation uncertainty is defined as the region the results of tests with systematic variations in each of the
containing 95% of the random error fluctuations squared. By four uncertainty sources considered in this paper, followed by
using the square, the random fluctuation uncertainty makes conclusions.
no assumption of the distribution shape and provides a 95%
confident uncertainty for asymmetric distributions. 2. Experimental setup
The instantaneous uncertainties from the uncertainty
surface are then propagated into the mean and variance or The facility used for this study was originally designed for
Reynolds normal stress as outlined in Wilson and Smith [27]. a different and unusual purpose, and therefore has some
Using this formulation, the propagated uncertainty on the mean unorthodox features. It was chosen for the present study
Uu is because it is capable of producing a laminar, high-Reynolds-

 2 number, channel flow, and therefore a jet with a laminar core
 1 N
Uu =  bu + pu2 , (2)
and very high shear. These features were desirable to provide
N i=1 i velocity fluctuation levels of similar magnitude to the PIV
random uncertainty.
where pu2 is the precision uncertainty on the mean value, N is
As shown in figure 1, the blower draws air from the lab
the total number of measurements acquired and the subscript i
environment and into the diffuser section. This is followed by
denotes the instantaneous uncertainty values obtained from
a flow control region made from open cell foam, which leads
the uncertainty surface (i.e. bu , ru+ , ru− and rû2 ) for each
into a contraction terminated by a short, straight channel. As
instantaneous realization.
the flow exits the channel to form a jet, an unsteady flow
Random errors will cause measurements to bias the
with regions of high shear surrounds a laminar potential core.
Reynolds normal stress making the bias uncertainty of the
The exit has height h = 9.5 mm and width of 8h. The
Reynolds normal stress one-sided and in the negative direction.
unique contraction is designed to provide a nearly uniform
Precision uncertainty impacts both the upper and lower
laminar profile at the jet exit for Reynolds numbers of up to
uncertainty band. According to Taylor [30]
 Re = 60 000 with approximately 70% of the jet profile within
pû2 = 1.96 2/(N − 1)s2u , (3) 0.5% of the maximum velocity.

4
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

The Reynolds number is defined as u/uy=0 u u / (uy=0 )2

ρ f umax h1/2 0.005
Re = (7) 1
μf
where the jet half-height h1/2 is the length scale, the velocity 0.004
scale (umax ) is the maximum velocity at the jet exit, ρ f is the
0.75
fluid density and μ f is the dynamic viscosity of the air. Two

u u / (uy=0 )2
Reynolds number cases were measured: Re = 2.57 × 104 and 0.003

u/uy=0
8.94 × 104 corresponding to umax = 10.1 and 35.1 m s−1 .
The cross-stream profile of the streamwise velocity was 0.5
measured by both hot wire and PIV. The jet and laboratory 0.002
environment were at the same atmospheric conditions: ρ =
1.0048 ± 0.00325 kg m−3 and μ = 1.88 × 10−5 ± 1.93 × 10−8 0.25
kg m s−1 . Atmospheric quantities were monitored during 0.001
acquisition of data to ensure similar conditions for both hot-
wire and PIV measurements.
0 0
Although acquired separately, both hot-wire and PIV 200 400 600 800 1,000 1,200
measurements were made in the same plane at the span-wise y (pixels)
center (z = 0) of the jet. Hot-wire measurements used a coated
single-sensor probe and a TSI IFA 100 Hot-Wire Anemometer Figure 2. The time-mean streamwise velocity and Reynolds normal
stress profiles of the jet. The jet exhibits regions of high shear at the
and were acquired using a 12-bit data A-D board. The sensor jet edges and uniform laminar flow at the jet core.
diameter was 3.5 μm, and the sensor was aligned with the jet
span. The sensor length was 1.5 mm and a response time of
12 μs was observed in the hot-wire signal under a square wave
test. The uncertainty of the hot-wire measurements is estimated
to be 2.5% of reading and stems from the calibration source
(calibration jet and pressure sensor) and the analogue to digital
data acquisition system. Precision uncertainty of the hot wire
was reduced below 0.001%. The hot wire was attached to a
stepper-motor-operated traverse (accurate within ±5 μm) that
was used to position the hot wire at discrete locations along the
jet profile. At each location, N = 262 144 measurements were
obtained and a sampling rate of f = 10 000 Hz. Jet profile
measurements were acquired at two streamwise locations: Figure 3. The locations of the instrumentation in the experimental
x/H = 0.1 and 1.25. In this manner, two accurate and resolved setup. The laser sheet and hot-wire probe are located in the same
plane with the camera positioned normal to the laser sheet.
cross-stream profiles of the flow were measured with different
magnitudes of shear. The seeder did not operate during hot- the sensitivity to both velocity components was demonstrated
wire measurements. to be the same within 5% (see appendix A). However,
PIV images were acquired in the same plane as the hot- PIV measures both velocity components individually and a
wire measurements (figure 3) using a 12-bit, 1376 × 1040 method of comparing velocity magnitudes and fluctuations
pixel imager intense CCD camera using a Nikon 105 mm is needed. While this is simple for the velocity magnitude
lens and a New Wave dual-cavity 50 mJ/pulse Nd:YAG laser (|u|2 = u2x + u2y ), the relation to combine the fluctuations
with attached sheet optics. This system was controlled with is not as straight forward. The relationship between the two
DaVis 7.2 from LaVision [29]. The flow was seeded using component fluctuations and the fluctuation magnitude is shown
oil droplets distributed at the blower inlet, far upstream of in appendix A to be
the flow conditioning. The flow rate of the secondary seeded  
u u = ux ux + uy uy , (8)
flow was less than 1% of the primary flow and negligible
differences were observed by a pitot probe with and without where ux ux and uy uy are the velocity fluctuations in the x and
seed. Droplets formed in this manner are typically assumed y directions and the fluctuations in the velocity magnitude are
to have a diameter near 1 μm. To obtain converged statistics, denoted by u u and are referred to herein as the Reynolds
N = 1000 images were acquired at an average of f = 3 Hz. normal stress.
The spatial resolution for these images was approximately The cross stream profile of the streamwise time-average
9 μm/pixel, while the laser sheet width was the same as the and Reynolds normal stress at the exit measured with the
hot-wire probe length (1.5 mm). hot wire are shown in figure 2. The core flow is laminar,
A single-sensor hot wire is sensitive to both the while the shear layers exhibit strong fluctuations. To detect
streamwise (ux ) and cross-stream (uy ) velocity components, the measurement fluctuations, ŭŭ, it is helpful to have minimal
and, if equally sensitive to both directions, the hot-wire output Reynolds normal stress, u u , in some regions.
is simply the magnitude of the velocity in the x and y One side of the channel terminates upstream of the other,
directions. Using a TSI hot-wire calibrator to pitch the probe, leading to non-symmetric velocity mean and Reynolds normal

5
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

y = 700 pixels
y = 250 pixels

10−4

Magnitude 10−5

10−6

10−7
101 102 10−2 10−1 100
f (Hz) λ (m)

Figure 4. Temporal spectra ( f ) and spatial spectra (λ) of the spatial scales found within the jet at the centerline (y = 700 pixels) and the
shear region (y = 250 pixels).

stress profiles in the shear layer. Noting that the PIV error spatial scale that can be resolved. Although this has a minimal
sources are the interest of this paper and not the jet profile, this effect on the measurement mean (except in regions of very high
asymmetry has no impact on the present study. shear), the measured variance and Reynolds normal stresses
PIV images were processed using two PIV algrothms: can be attenuated as a result of spatial filtering.
DaVis 7.2 [31] and PRANA [26, 28]. Data computed using the Velocity scales within the jet are observed in the temporal
DaVis algorithm were processed with square interrogations spectra measured by the hot wire which are shown in
regions with an initial window size of 32 × 32 and with two figure 4. Using Taylor’s frozen-turbulence hypothesis, the local
consecutive passes at 16×16 using a standard cross-correlation mean velocity is used to convert the temporal spectra into
(SCC) algorithm. All interrogation regions were overlapped by spatial scales, allowing an estimate of the required temporal
50%. Results were then post-processed using four parameters and spatial resolution. The velocity spatial and temporal
[29]: an allowable pixel range (vectors displacing more than scales within this jet are large and drop-off quickly into the
15 pixels are rejected), correlation peak ratio, neighboring measurement noise at small scales (λ  10–50 mm and
vectors median filter and small groups (spurious vectors f  50–100 Hz, respectively). We conclude that the velocity
in groups smaller than 5 vectors are thrown out). Image scales in the flow are much larger than the resolution of either
processing from the PRANA algorithm used the same settings measurements, and we anticipate minimal spatial filtering.
used in [2] for the robust phase correlation (RPC) method
[26, 28]. Both DaVis and PRANA results are obtained for
interrogation regions of hIA = 16 pixels across. 3. Results
Given the accuracy of the hot-wire traverse, images of the
The effects of each error source [i.e. particle image density
hot-wire sensor at several known locations were acquired and
used to calibrate the PIV images. Profiles from the PIV data (dρ ), diameter (dτ ), and displacement (x) and velocity
were extracted at the hot-wire profile locations (as measured gradients] on the measured velocity mean and Reynolds
from the hot-wire images) for direct comparison to the hot- stresses will be demonstrated. It is possible to systematically
wire results. vary dρ , dτ and x experimentally. Since the center of the jet
PIV results are given in dimensions of pixels for does not exhibit shear, the impact of these three error sources
particle displacement (velocity) and spatial coordinates and will be observed there. The parameter space values of error
pixels/pixel for gradients to remain consistent with [2]. All hot- sources are given in table B1. When one error source is varied,
wire velocities are also converted to pixels using the PIV spatial the other two are held fixed at a nominal, optimal value. These
calibration constant and time interval between images, t, to measured nominal and varying error source values are given
allow for direct comparisons. Since the hot-wire uncertainty in appendix B.
was small, and since the sensor is well suited to the flow To improve clarity, only the left side of the profile (the
in question, the hot-wire measurements will be considered lower side of the jet in figure 1) will be shown. Most values
‘correct’ when interpreting the PIV results. are normalized by the time-averaged centerline velocity as
measured by the hot wire, uy=0 .
The automatic uncertainty algorithm by Timmins et al
2.1. Spatial and temporal resolution
[2] is indifferent to regions of multiple error sources and
Spatial-filtering is a concern for both the hot-wire and will account for the combination of these error sources at
PIV measurements. The PIV measurement can be filtered all locations. Individual error source effects (particle image
by averaging through the sheet thickness and across the density, diameter and displacement) are observed at the
interrogation region. The hot-wire measurements are averaged centerline, while these combined with gradient effects are
over the probe length. In each case, this limits the minimum observed at the jet shear layer.

6
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

PRANA DaVis Hot Wire

0.5 pixels 0.7 pixels 0.9 pixels

u/uy=0 1.5

0.5

1.1 pixels 2.1 pixels 3.2 pixels

1.5
u/uy=0

0.5

0
200 400 600 200 400 600 200 400 600
x (pixels) x (pixels) x (pixels)
Figure 5. Profiles of the time-averaged velocity magnitude, u, from the hot-wire and PIV measurements (DaVis and PRANA) with 16 × 16
pixel interrogation regions for varying small displacements.

We seek to estimate the uncertainty of the time average lower limit is not observed. We note that while it is simple and
velocity and the Reynolds normal stress. Therefore, the data advisable to ensure that the largest velocity region in a PIV
are presented as u u with an uncertainty band rather than ûû. data set experiences a large displacement, for many flow of
interest (e.g. jets), there will always be regions of slow flow,
3.1. Effects of particle image displacement which will experience large uncertainties like those shown in
figure 5.
The time interval t between consecutive images controls While the small displacement cases begin to converge on
the particle image displacement x at all locations in an image the hot-wire mean velocities as the displacement is increased,
set. We divide our result into two groups: small and large Reynolds normal stresses are dramatically elevated for all
displacements. The dominant source of error from particle small displacement cases. As these errors are too large to
image displacement is from the PIV algorithm’s limited ability clearly plot, they are only shown for the large displacement
to resolve sub-pixel displacement. Consequently, significant cases (cf figure 6).
bias and random errors are observed for displacements on the For all large displacement cases, the time-averaged
order of one pixel (Nogueira et al [32]). velocity magnitude agrees with the hot-wire result within
As t, and thus x, is increased, the dynamic range of the hot-wire uncertainty. However, significant error can still
the measurements is improved at the same rate. Therefore, be seen in the Reynolds normal stress except when the
the error from sub-pixel displacement decreases relative to displacements are very large, as shown in figure 6. As
x. This is clearly evident in the small displacement cases the displacement is increased, the Reynolds normal stress
(figure 5) as sub-pixel errors generate a strong bias on the converges to the value obtained using the hot wire. While
time-averaged velocity, which are reduced as the dynamic particle image displacements of x ≈ dτ are sufficient to
range is improved. Uncertainty levels for DaVis and PRANA provide accurate time-averaged velocity measurements when
are denoted with uncertainty bands. The bias is accurately sufficient data are acquired, accurate Reynolds normal stress
estimated by the measured mean uncertainty for the PRANA (and presumably, the shear stress) measurements require much
case but is under-estimated for the DaVis case. Timmins et al larger particle image displacements (x  4dτ ) for both the
[1, 2] suggest using a lower limit for the uncertainty to account PRANA and DaVis algorithms.
for additional, small error sources that are not accounted for by Uncertainty levels for DaVis and PRANA are denoted
the uncertainty surface, and suggested a value of 0.02 pixels with dashed lines in figure 6. The hot-wire results provided
for the PRANA RPC and a much larger value for SCC. A reference values, and therefore the uncertainties should
lower uncertainty limit is not used for the DaVis results and account for the difference between hot-wire and PIV results.
may account for the underestimated bias. If a lower limit of While the shear layer has little impact on the PIV measurement
0.1 pixels were used for the DaVis results, complete coverage of the time-averaged velocity, elevated Reynolds normal
is observed for the mean velocity uncertainty. For all PRANA stress levels are observed there. The PRANA uncertainty at
cases, the estimated uncertainties are the same magnitude or these locations, however, still accounts for the correlated,
larger than the lower uncertainty limit and the effect of the combined error effects. The uncertainty surface method

7
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

PRANA DaVis Hot Wire

UPRANA UDaVis

1.1 pixels 2.1 pixels 3.2 pixels

u u / (uy=0 )2 0.005

0.0025

0
4.3 pixels 5.3 pixels 6.4 pixels
0.005
u u / (uy=0 )2

0.0025

0
8.6 pixels 10.7 pixels 15.9 pixels
0.005
u u / (uy=0 )2

0.0025

0
200 400 600 200 400 600 200 400 600

x (pixels) x (pixels) x (pixels)

Figure 6. Profile of the measured Reynolds normal stress u u and uncertainty from the hot-wire measurement and the PIV measurement
using the DaVis and PRANA algorithms with 16 × 16 pixel interrogation regions for varying large particle image displacements.

(a) (b)

10−2
u u / (uy=0 )2

Hot Wire
PRANA
DaVis

10−3

10−4
0 5 10 15 20 0 5 10 15 20

Δx (pixels) Δx (pixels)
Figure 7. Reynolds normal stress, u u , from hot-wire, DaVis and PRANA algorithm measurements for varying particle displacements, x
at (a) the shear region and (b) the jet centerline. The solid line denotes the value from the hot-wire measurement.

properly accounts for the measurement error on the Reynolds figure 7, the Reynolds normal stress is shown in (a) the shear
normal stress due to the combined error sources for the region and (b) the jet core. The errors on the normalized mean
PRANA results in the jet core as well as the shear layer. The velocity are a strong function of particle image displacement
DaVis uncertainty under-estimates the measurement error for for both the PRANA and DaVis algorithms. As the dynamic
small particle image displacements. range increases, the magnitude of the measurement errors
Along the jet centerline, no shear exists and the error is decrease relative to the total particle image displacement
due to particle image density, diameter and displacement. In of the measurement. Elevated Reynolds normal stress for
the shear layer, the error depends on all four parameters. In the x = 15.9 pixel case using DaVis 7.2 result from a

8
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

Figure 8. Examples of PIV images for various particle image diameter values: dτ = 1.9, 2.2, 2.5, 3.1, 4.2 and 5.8 pixels.

particle image diameter above the nominal value for this 8

case (see appendix B), which DaVis is more sensitive to than Real Particle Diameter
PRANA. Ideal Particle Diameter
In the shear layer, the effects of flow gradients are
combined with the effects of the other three error sources. In 6
this region, errors are dominated by flow gradient for particle

dτ (pixels)
image displacements larger than x > 1 pixel. A trend similar
to the jet centerline is observed at sub-pixel displacements
4
indicating dominance by the particle image displacement
errors for these cases. Uncertainties are accurately quantified in
the shear layer and for most large particle image displacement
cases (x > 3 pixels) at the jet centerline. We note that the 2
method is better able to quantify uncertainties when they are
large than when they are very small.

0
3.2. Effects of particle image diameter 0 5 10 15 20 25 30 35
An easily-controlled contributor to PIV measurement error is f
particle image diameter dτ . Particle image diameter is strongly Figure 9. Comparison of actual particle image diameters and ideal
affected by the image diffraction through the lens aperture. diffraction-limited particle image diameters as a function of the f # .
With an ideal lens, the 2σ (95% width) particle image diameter
should vary with the f # through the equation [3]
Table 1. The lens f # and the resultant particle image diameter in
dτ = M 2 d 2p + ds2 , (9) pixels. For each case, the mean I and standard deviation sI of
intensities of the entire image for datasets are also given in units
where M denotes the magnification, d p is the actual particle counts out of a maximum of 4096.
diameter and ds is the diffraction-limited spot diameter,
f# dτ I sI
ds = 2.44 (1 + M) f λ,#
(10) 5.6 1.9 53.8 14.14
where λ is the wavelength of the light sheet, and f # is the ratio 8 2.2 52.7 13.42
11 2.5 52.5 14.34
of the lens focal length, f , and aperture diameter, Da . Given 16 3.1 53.3 11.24
the very small particle diameter, the first term under the radical 22 4.2 54.9 11.24
in equation (9) is dominated by the second, and the particle 32 5.8 57.7 11.08
image size is not a function of the particle size.
Particle image diameter was controlled by varying the
lens f # . The particle image densities and intensities were and lens aberrations) is directly related to the peak width of an
autocorrelation of the image and can be estimated from [3]
held nominally constant through modifications in laser power √
and seed quantities for all datasets. A reference image at the dτ = 2 2σcorr (11)
smallest f # (2.8) was acquired and densities and intensities where σcorr is the autocorrelation peak width and dτ is the e−2
were visually matched for subsequent image sets. Examples particle image diameter. The autocorrelation peak width σcorr
of the density and intensity for a 64×64 pixel region are shown is calculated from the standard deviation of a Gaussian fit of
in figure 8 for several particle image diameters from dτ = 1.9 the autocorrelation peak.
to 5.8 pixels. Verification of constant seed density and image For this study, local particle image diameters are estimated
intensity was achieved by monitoring the mean I and standard using equation (11) for autocorrelated 128 × 128 pixel
deviation sI of intensities over the entire image (shown in interrogation regions. The actual particle image diameters
table 1). are compared to ideal particle image diameters in figure 9.
For small f # (large apertures), typical lenses experience Deviation from the ideal particle at small f # are attributed to
aberrations which increase the particle image diameter. lens aberrations. Using this particular experimental setup, the
However, the actual particle image diameter as seen by a minimum particle image diameter corresponds to dτ = 1.9
camera sensor (including ‘true’ particle size, image diffraction pixels ( f #  5.6).

9
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

PRANA DaVis Hot Wire

UPRANA UDaVis

1.9 pixels 2.2 pixels 2.5 pixels

0.005
u u / (uy=0 )2

0.0025

0
3.1 pixels 4.2 pixels 5.8 pixels
0.005
u u / (uy=0 )2

! Outside
Surface
0.0025

0
200 400 600 200 400 600 200 400 600

x (pixels) x (pixels) x (pixels)

Figure 10. Cross-stream profiles of the measured Reynolds normal stress u u using the DaVis and PRANA algorithms with 16 × 16 pixel
interrogation regions for varying particle image diameters. The dτ = 5.8 pixels case is outside the uncertainty surface (denoted by ‘! Outside
Surface’) and uses the uncertainty for a particle image diameter of dτ = 5.0 pixels.

Random error will increase as particle image diameter Table 2. Mean and standard deviation of intensities of the entire
increases beyond 3 pixels [2, 22]. As the particle image image along with relative intensity and density increases relative to
the most dense case for datasets used for the particle image density
diameter is varied, negligible differences are observed on the
error source case.
time-averaged velocity profiles (not shown). This, however,
is not the case for the Reynolds normal stress shown in Particle image Intensity
density dρ % Change mean % Change
figure 10. For most of the cases, the PIV-reported Reynolds
normal stress is above the hot-wire ‘true’ value. Uncertainty 0.002 3 55.37 12
estimates correctly account for the measurement errors in most 0.012 17 89.25 19
of the cases. For the largest particle image diameter (dτ = 0.030 43 173.7 37
0.047 68 285.4 61
5.8 pixels), the average particle image diameter is outside 0.058 84 389.5 84
the uncertainty surface from Timmins et al [1, 2] and uses 0.069 100 465.5 100
the uncertainty for a particle image diameter of dτ = 5.0
pixels. When the error sources are within the uncertainty
surface generated by Timmins et al [1, 2], the uncertainty
surface is capable of accounting for fluctuation measurement
the autocorrelation Rh , average particle intensity Ip and particle
error in both the jet core and shear layer regions for PRANA.
DaVis uncertainties are often under-estimated at all centerline image diameter dτ :
locations, but accurate at the combined shear regions where Rh
dρ = 14.8 − 3.3, (12)
the lack of a minimum uncertainty value has no impact. dτ Ip A
Prasad [33], Westerweel [34], and Adrian and Westerweel
[3] demonstrate optimal particle image diameters at 2.0  with the interrogation domain size denoted by A. The basis of
dτ  2.5 pixels. In this study the minimum error is observed this relationship is the subject of a future publication, but the
for particle image diameters near dτ = 2.2 and 3.1 pixels. In resultant values were verified through manual particle counting
the shear layer region, the shear strongly dominates measured for all datasets. Particle image density was varied by increasing
Reynolds normal stress errors. the seed particles within the experimental domain. Examples
of these images are shown for a 64 × 64 pixel region in
3.3. Effects of particle image density figure 11. To give confidence that this experiment was
controlled as desired, the computed particle density and
Particle image density is defined as the average number of the image mean intensity relative to the most dense case
visible particles within each interrogation region. The particle are reported in table 2. These data showed that as the
image density is estimated within the automatic uncertainty particle density increased, the mean image intensity increased
algorithm using an empirical fit based on the peak height of at the same rate.

10
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

Figure 11. Examples of PIV images for given particle image density error sources.

PRANA DaVis Hot Wire

UPRANA UDaVis

0.002 part./pix.2 0.012 part./pix.2 0.030 part./pix.2

0.005
u u / (uy=0 )2

! Outside
Surface
0.0025

0.047 part./pix.2 0.058 part./pix.2 0.069 part./pix.2

0.005
u u / (uy=0 )2

! Outside ! Outside
Surface Surface
0.0025

0
200 400 600 200 400 600 200 400 600

x (pixels) x (pixels) x (pixels)

Figure 12. Cross-stream profiles of the measured Reynolds normal stress u u using the DaVis and PRANA algorithms with 16 × 16 pixel
interrogation regions for varying particle image densities. The dρ ≈ 0.002, 0.058 and 0.069 particles/pixel2 cases were outside the
uncertainty surface (denoted by ‘! Outside Surface’) and use the nearest uncertainty surface estimate.

The effects of particle image density on PIV demonstrated during the previous sections by significantly
measurements are studied using particle image densities of elevated fluctuation levels in the shear layer. Although shear
dρ ≈ 0.002, 0.012, 0.030, 0.047, 0.058 and 0.069 particles/ will introduce error to all PIV algorithms, the magnitude
pixel2 . For the most dense case (dρ  0.058 particles/pixel2 ), of measurement error will likely differ between algorithms.
particle images overlap. Flow gradients introduce error into a PIV measurement in two
Both time-averaged velocities and Reynolds normal ways: by broadening the correlation peak and by introducing
stresses are negligibly affected by particle image density random, instantaneous errors from non-uniform particle image
for most cases (figure 12); however, at low particle image densities.
densities, insufficient seeding introduces measurement noise, Difficulties in measuring instantaneous, local velocity
especially in regions of shear [2]. Reynolds normal stress gradients often motivate the use of the time-averaged
measurement errors and uncertainties are larger for the velocity gradients to assess gradients within noisy PIV
dρ ≈ 0.047 particles/pixel2 case. While the individual effects measurements. Finite differencing methods amplify noise
of particle image density on measurement error at the jet
of the instantaneous, local gradient calculations; however,
centerline are often negligible, the correlated effects of density
local mean gradients based off the time-averaged velocity
and gradients are evident near the shear layer, as evidenced by
profile tend to under-predict instantaneous, local velocity
elevated Reynolds normal stresses as the particle image density
gradients. Consequently, time-averaged gradients under-
decreases.
estimate measurement uncertainties due to the nonlinear
relationship between measurement error and gradients. This
3.4. Effects of shear
is of particular interest in turbulent flows, where a zero
Of the error sources studied by Timmins et al [2], the mean gradient may exist while the flow exhibits large
largest uncertainties were found to be due to instantaneous local, instantaneous gradients. The instantaneous, local
velocity gradients (or shear) in the flow. This is also gradient measurement uncertainties in this study are estimated

11
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

PRANA DaVis Hot Wire

UPRANA UDaVis

0.07 pixels/pix. 0.07 pixels/pix.

1.5 0.005

u u / (uy=0 )2
u/uy=0
1
0.0025
0.5

0 0

0.09 pixels/pix. 0.09 pixels/pix.

1.5 0.02

u u / (uy=0 )2
u/uy=0

1
0.01
0.5

0 0

0.14 pixels/pix. 0.14 pixels/pix.

1.5 0.015

u u / (uy=0 )2
u/uy=0

1 0.01

0.5 0.005

0 0
200 300 400 200 300 400

x (pixels) x (pixels)

Figure 13. Cross stream profiles of the time-averaged velocity magnitude, u, and Reynolds normal stress, u u , for gradients of
du/dy ≈ 0.07, 0.09 and 0.14 pixels/pixel from the hot-wire, DaVis and PRANA measurements. Uncertainty bands indicate the uncertainty
on u, while dashed lines indicate the uncertainty on u u .

automatically using the high accuracy, low noise gradient and PRANA algorithms begin to deviate further from the hot-
estimator introduced by Karri et al [35]. wire results. Uncertainty estimates on the Reynolds normal
Data were acquired for three different cases in which the stress caused by the flow gradients show the Reynolds normal
time-averaged gradient levels in the shear layer were observed stress uncertainty Uu− u from the PIV results are composed
to be approximately du/dy ≈ 0.07, 0.09 and 0.14 pixels/pixel. primarily of uncertainties from gradients within the flow.
The two largest shear cases were acquired at two streamwise The du/dy ≈ 0.09 pixels/pixel case, which was acquired
locations of the jet with a maximum time-averaged velocity of farther downstream, introduces an interesting phenomenon:
umax = 35.1 m s−1 , while the smallest shear case corresponds increased measurement error with decreased flow gradients.
to umax = 10.1 m s−1 . While this cannot be confirmed, we believe this is generated
Hot-wire, DaVis and PRANA profiles of the time- by an increase in through-plane motion, which degrades the
averaged velocity u are shown in figure 13 for the three correlation signal strength and is not accounted for by the
shear levels. For all cases, accurate measurements of the uncertainty surface.
time-averaged velocity magnitude are obtained. Velocities at For all cases, the difference in PIV measured Reynolds
the jet centerline compare well between hot-wire and PIV normal stress from time-averaged measured fluctuation
measurements. In the high shear region, the PIV time-averaged uncertainty Uu− u agrees well with the hot-wire results for
velocity profiles tend to be smeared or smoothed due to the PRANA algorithm. In all but the large gradient case,
averaging over the interrogation region. This effect, however, the uncertainty for the DaVis algorithm significantly under-
is small and has little impact on the mean profiles. predicts the fluctuation error.
Profiles of the Reynolds normal stress u u for all three Figure 13 demonstrates the potential for different
shear cases are also shown in figure 13. These profiles consist dominant error sources from different algorithms. For the
of a core region with no shear and small velocity fluctuations du/dy = 0.07 and 0.14 pixels/pixel cases, the flow gradients
and the shear layer regions that contain varying levels of shear. are strongly one-dimensional, reducing the fluctuation error
In the shear layer region, large gradients cause elevated levels observed from DaVis; however, for the downstream case
of Reynolds normal stress. As the gradient increases, the DaVis (du/dy = 0.09 pixels/pixel) vortical structures begin to

12
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

occur and flow gradients are no longer dominant in the support of the DOE through NEUP grant number 00118627
y-direction, increasing measured fluctuation levels from and Idaho National Laboratory LDRD Project NE-156 is
DaVis. The opposite trend is observed in the PRANA results gratefully acknowledged.
and is attributed to the lack of window deformation within
the version of PRANA used by Timmins et al [2]. Potential
Appendix A. Magnitudes of fluctuations
additional error sources for these cases include out-of-plane
motion, which is difficult to detect and was therefore not
A single hot wire is sensitive to both the streamwise (ux )
included in the uncertainty surface method. This may cause
and spanwise (uy ) velocity components, and therefore its
the uncertainty surface to under-predict ‘true’ errors.
output is essentially the magnitude of the velocity in the x
and y directions. We assume that uz is very small in all of
4. Conclusion the cases presented here and that therefore u2 = u2x + u2y .
This relationship is used to convert the PIV data into results
A previously published method for determining the uncertainty comparable to the hot-wire measurements. The hot-wire
stemming from the vector field computation in PIV probe must be equally sensitive to both components for
measurements was demonstrated on real PIV data. The this conversion to be accurate. The response of the single
‘uncertainty surface method’ uses the response of a PIV code to wire probe to velocity components normal to each other was
various error sources coupled with information about the local, assessed using a TSI hot-wire calibrator that allows the probe
instantaneous values of these sources to find an uncertainty to be rotated while maintaining the sensor position. This test
for each vector. The four error sources were: (1) particle yielded a difference in normal velocity components within 5%
displacement, (2) particle image density, (3) particle image
for a probe rotation of 90◦ . We can write that, for the hot-wire
diameter and (4) velocity gradients. A major advantage of this
sensor,
method is the ability to account for any combination of the
2
four sources. (u + u )2 = ux + ux + (uy + uy )2 , (A.1)
The flow field of a planar jet was measured with PIV and
hot-wire anemometry. The effects of PIV uncertainty on the
time-average velocity and the Reynolds normal stresses, were 2 
u + 2uu + u u
also assessed.      
Random uncertainty from these four sources affected = u2x + u2y + ux ux + uy uy + 2ux ux + 2uy uy . (A.2)
the rate at which converged statistics were acquired. These
errors are two-sided and, when sufficient data are obtained, a An informative way to consider these equations is that the
converged mean that agrees with the hot-wire solution was left-hand side is what is measured by the hot wire while the
measured. However, the measurement noise increased the right-hand side can be obtained from the PIV measurement.
Reynolds normal stress reported by the PIV measurement The time-average of the cross terms is zero. This provides
system resulting in Reynolds normal stress levels that were a root-mean-square relationship for the mean velocity and
overestimated. Of the four error sources, velocity gradients fluctuations components,
were the largest contributor to PIV random uncertainty.    
For pixel displacements of x  dτ , the measured mean u2 + u u = u2x + u2y + ux ux + uy uy . (A.3)
velocity was strongly biased, while Reynolds normal stresses
Using this relationship, a direct comparison of PIV and hot-
were significantly overestimated for x  4dτ . This is due
wire measurements can be made. For simplicity, all variables
to a combination of decreased dynamic range for the smaller
discussed throughout will be time-averaged values. This gives
pixel displacements and sub-pixel interpolation uncertainties.
the magnitude statistics the relations,
A previously determined minimum uncertainty for the  
PRANA PIV algorithm was used throughout the paper. No u2 = u2x + u2y , (A.4)
such value has yet been determined for the DaVis algorithm,
 
and as a result, uncertainties did not cover the ‘true’ results in u u = ux ux + uy uy . (A.5)
several cases where the uncertainty values were small.
The work presented does not account for through-plane Using this form, PIV data and hot-wire data are directly
motion, and some evidence was presented that as random comparable.
through-plane motion increases, the performance of the
method degrades. In principle, the uncertainty surface method Appendix B. Nominal values of error source datasets
could be extended to include these effects.
The magnitudes of the error sources are measured using
Acknowledgments the methods described previously. The measured values
for particle image density, diameter and displacement and
The authors would like to thank Professor John Foss for flow gradients are used to estimate the instantaneous,
helpful discussions on interpretation of the hot-wire signals local uncertainty on the measurement using the automatic
used in this work. We also thank Scott Warner for providing uncertainty method proposed by Timmins et al [1, 2]. Error
equation (12) for estimation of particle image density. The source values for all PIV datasets are given in table B1.

13
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

Table B1. The nominal and varying error source values of error [12] Nobach H and Bodenschatz E 2009 Limitations of accuracy in
source for all datasets. PIV due to individual variations of particle image intensities
Exp. Fluids 47 27–38
Error dρ , particles/ dτ , x, Maximum gradient
[13] Kaehler C J, Scharnowski S and Cierpka C 2012 On the
source pixels2 pixels pixels (du/dy)
uncertainty of digital PIV and PTV near walls Exp. Fluids
dρ 0.002 2.6 10.7 0.07 52 1641–56
dρ 0.012 3.0 10.7 0.07 [14] Adrian R J 1997 Dynamic ranges of velocity and spatial
dρ 0.030 3.0 10.7 0.07 resolution of particle image velocimetry Meas. Sci. Technol.
dρ 0.047 2.9 10.7 0.07 8 1393–8
dρ 0.058 2.9 10.7 0.07 [15] Westerweel J, Dabiri D and Gharib M 1997 The effect of a
dρ 0.069 2.9 10.7 0.07 discrete window offset on the accuracy of cross-correlation
analysis of digital PIV recordings Exp. Fluids 23 20–8
dτ 0.010 1.9 10.7 0.07 [16] Persoons T and O’Donovan T S 2011 High dynamic velocity
dτ 0.012 2.2 10.7 0.07 range particle image velocimetry using multiple pulse
dτ 0.012 2.5 10.7 0.07 separation imaging Sensors 11 1–18
dτ 0.010 3.1 10.7 0.07
[17] Rabinovich S G 2005 Measurement Errors and Uncertainties
dτ 0.007 4.2 10.7 0.07
3rd edn (Berlin: Springer)
dτ 0.007 5.8 10.7 0.07
[18] Carr Z R, Ahmed K A and Forliti D J 2009 Spatially correlated
x 0.032 3.0 0.5 0.07 precision error in digital particle image velocimetry
x 0.032 3.1 0.7 0.07 measurements of turbulent flows Exp. Fluids 47 95–106
x 0.034 3.1 0.9 0.07 [19] Gui L, Longo J and Stern F 2001 Towing tank PIV
x 0.033 3.1 1.1 0.07 measurement system, data and uncertainty assessment for
x 0.031 3.1 2.2 0.07 DTMB Model 5512 Exp. Fluids 31 336–46
x 0.033 3.0 3.2 0.07 [20] Adeyinka O B and Naterer G F 2005 Experimental uncertainty
x 0.032 3.0 4.3 0.07 of measured entropy production with pulsed laser PIV and
x 0.030 3.0 5.3 0.07 planar laser induced fluorescence Int. J. Heat Mass Transfer
x 0.031 3.0 6.4 0.07 48 1450
x 0.031 3.0 8.6 0.07 Adeyinka O B and Naterer G F 2005 Experimental uncertainty
x 0.032 3.0 10.7 0.07 of measured entropy production with pulsed laser PIV and
x 0.033 3.3 15.9 0.07 planar laser induced fluorescence (vol 48, pg 1450, 2005)
du/dy 0.031 3.0 6.4 0.07 Int. J. Heat Mass Transfer 48 3889 (corrigendum)
du/dy 0.025 2.7 7.0 0.09 [21] Lazar E, DeBlauw B, Glumac N, Dutton C and Elliott G 2010
du/dy 0.025 2.7 7.0 0.14 A practical approach to PIV uncertainty analysis 27th AIAA
Aerodynamic Measurement Technology and Ground Testing
Conf. number AIAA 2010-4355 (Chicago, IL)
References [22] Raffel M, Willert C and Kompenhans J 1998 Particle Image
Velocimetry (Berlin: Springer)
[23] Stanislas M, Okamoto K and Kaehler C 2003 Main results of
[1] Timmins B H, Smith B L and Vlachos P P 2010 Automatic the first international PIV challenge Meas. Sci. Technol.
particle image velocimetry uncertainty quantification Proc. 14 R63–89
ASME Fluids Engineering Division Summer Meeting Paper
[24] Stanislas M, Okamoto K, Kahler C J and Westerweel J 2005
number FEDSM-ICNMM2010-30724
Main results of the second international PIV challenge Exp.
[2] Timmins B H, Wilson B M, Smith B L and Vlachos P P 2012
Fluids 39 170–91
A method for automatic estimation of instantaneous local
uncertainty in particle image velocimetry measurements [25] Charonko J J and Vlachos P P 2012 Estimation of uncertainty
Exp. Fluids 53 1133–47 bounds from cross correlation peak ratio for individual PIV
measurements ASME Fluids Engineering Summer Meeting
[3] Adrian R J and Westerweel J 2011 Particle Image Velocimetry
Paper FEDSM2012-72475
(New York: Cambridge University Press)
[4] Stanislas M, Okamoto K, Kaehler C J, Westerweel J [26] Eckstein A C and Vlachos P P 2009 Digital particle image
and Scarano F 2008 Main results of the third international velocimetry (DPIV) robust phase correlation Meas. Sci.
PIV challenge Exp. Fluids 45 27–71 Technol. 20 055401
[5] Keane R D and Adrian R J 1992 Theory of cross-correlation [27] Wilson B M and Smith B L 2013 Taylor-series and
analysis of PIV images Appl. Sci. Res. 49 191–215 Monte-Carlo-method uncertainty estimation of the width of
[6] Huang H, Dabiri D and Gharib M 1997 On errors of digital a probability distribution based on varying bias and random
particle image velocimetry Meas. Sci. Technol. 8 1427–40 error Meas. Sci. Technol. 24 035301
[7] Gui L and Wereley S T 2002 A correlation-based continuous [28] Eckstein A C and Vlachos P P 2009 Assessment of advanced
window-shift technique to reduce the peak-locking effect in windowing techniques for digital particle image
digital PIV image evaluation Exp. Fluids 32 506–17 velocimetry (DPIV) Meas. Sci. Technol. 20 075402
[8] Nogueira J, Lecuona A, Nauri S, Legrand M [29] LaVision 2006 Davis 7.2 LaVision GmbH Göttingen,
and Rodriguez P A 2011 Quantitative evaluation of PIV Germany
peak locking through a multiple t strategy: relevance of [30] Taylor J R 1997 An Introduction to Error Analysis: The Study
the rms component Exp. Fluids 51 785–93 of Uncertainties in Physical Measurements (Sausalito, CA:
[9] Westerweel J 2008 On velocity gradients in PIV interrogation University Science Books)
Exp. Fluids 44 831–42 [31] LaVision Inc. 301 W Michigan Ave, Suite 403, Ypsilanti, MI
[10] Scarano F 2002 Iterative image deformation methods in PIV 48197, USA
Meas. Sci. Technol. 13 R1–19 [32] Nogueira J, Lecuona A, Nauri S, Legrand M
[11] Roth G I and Katz J 2001 Five techniques for increasing the and Rodriguez P A 2009 Multiple t strategy for particle
speed and accuracy of PIV interrogation Meas. Sci. Technol. image velocimetry (PIV) error correction, applied to a hot
12 238–45 propulsive jet Meas. Sci. Technol. 20 074001

14
Meas. Sci. Technol. 24 (2013) 035302 B M Wilson and B L Smith

[33] Prasad A K, Adrian R J, Landreth C C and Offut P W 1992 [35] Karri S, Charonko J J and Vlachos P P 2009 Robust wall
Effect of resolution on the speed and accuracy of particle gradient estimation using radial basis functions and proper
image velocimetry interrogation Exp. Fluids 13 105–16 orthogonal decomposition (POD) for particle image
[34] Westerweel J 1997 Fundamentals of digital particle image velocimetry (PIV) measured fields Meas. Sci. Technol.
velocimetry Meas. Sci. Technol. 8 1379–92 20 045401

15