Sie sind auf Seite 1von 15

Experiments in Fluids 38 (2005) 146160

DOI 10.1007/s00348-004-0873-4

A comparison between snapshot POD analysis of PIV velocity


and vorticity data
J. Kostas, J. Soria, M. S. Chong

146
Abstract Proper orthogonal decomposition (POD) was not, hidden amongst the incoherent turbulent motions. In
performed on both the fluctuating velocity and vorticity addition, the volume of data can be so large that it must be
fields of a backward-facing step (BFS) flow at Reynolds summarised in a concise manner so that the most useful
numbers of 580 and 4,660. The data was obtained from information about the physical processes may be
particle image velocimetry (PIV) measurements. The extracted.
vorticity decomposition captured the fluctuating enstro- Proper orthogonal decomposition (POD) has proved
phy more efficiently than the equivalent velocity field to be an effective method for identifying dominant fea-
decomposition for a given number of modes. Coherent tures and events in experimental and numerical data. It
structures in the flow are also more easily identifiable has the added advantage of effectively compressing or
using vorticity-based POD. A common structure of the summarising large quantities of data so that the most
low-order vorticity POD modes suggests that a large-scale useful information about the physical processes occur-
similarity, independent of the Reynolds number, may be ring may be extracted. It was first introduced in the
present for the BFS flow. The POD modes obtained from a context of turbulence by Lumley (1967), but was sug-
vorticity-based decomposition would help in determining gested independently by several researchers before this,
a basis for constructing simplified vortex skeletons and including Kosambi (1943), Loe`ve (1945), Karhunen
low-order flow descriptions based on the vorticity of tur- (1946), Pougachev (1953) and Obukhov (1954) (see
bulent flows. Berkooz et al. 1993 for more details on these references).
POD also goes by the names Karhunen-Loe`ve decompo-
sition, principal component analysis (PCA), singular
1 value decomposition (SVD) and empirical eigenfunction
Introduction decomposition, amongst others. Its uses vary: random
variable analysis, image processing, signal analysis, data
1.1 compression, process identification and control in
Proper orthogonal decomposition (POD) chemical engineering, oceanography, even psychology
It is nowadays accepted that turbulent flows are charac- and economics (see Berkooz et al. (1993) for references
terised by organised motions, commonly referred to as on each discipline).
coherent structures, within which most of the essential When performing a POD of a flow, we are looking for
flow physics is believed to be buried. Understanding their fluid motions that contribute most to the energy of the
dynamics and interactions with other coherent structures flow (defined as the mean square fluctuating value of the
in the flow will lead to an increased understanding of flow variable under investigation). The decomposition
turbulence and its control. The process of identifying flow (akin to a Fourier series decomposition) results in a set of
patterns in experimental and computer-generated data is modes that represent an average spatial description of
by no means a trivial task, since they are, more often than structures containing most of the energy and which are
predominantly associated with large-scale structures. They
do not necessarily have to correspond to coherent struc-
Received: 8 October 2003 / Accepted: 15 August 2004 tures. The modes may also correspond to events in the
Published online: 12 January 2005 flow that contribute the most, in a statistical sense, to the
 Springer-Verlag 2005 energy of the flow.
In recent years, POD has begun to gain widespread use
J. Kostas, J. Soria (&) as a technique for identifying not only the dominant fea-
Laboratory for Turbulence Research in Aerospace and Combustion,
Mechanical Engineering Department, Monash University,
tures in flowscoherent structuresbut also for con-
3800 Monash, Victoria, Australia structing low-dimensional models that describe the
E-mail: julio.soria@eng.monash.edu.au principal flow dynamics with the least number of modes
M. S. Chong
(Bonnet 1988; Kevlahan 1994). A POD analysis requires
Mechanical and Manufacturing Engineering Department, the knowledge of the two-point spatial correlation func-
Melbourne University, 3052 Melbourne, tion, which can be found using multi-point measurements
Victoria, Australia such as rakes of hot-wire probes, scalar visualisation data,
The financial support of the Australian Research Council (ARC) particle image velocimetry (PIV) data and 2-D and 3-D
is greatly appreciated. direct numerical simulation data. Naturally, using POD on
3-D data is the preferred choice for a full understanding of components per vector field, the snapshot POD technique
turbulent flow physics. was employed to reduce the computational effort involved.
Despite its usefulness in the extraction of dominant Unfortunately, the sampling period, 10 s, was insufficient
flow structures, the application of POD has limitations. to ensure that each set of data was stationary in the mean,
The less deterministic a flow is, the less efficiently it will be but this was unavoidable due to limited computer re-
represented by the POD expansion. For example, the ki- sources. Nevertheless, Huang (1994) found 75% of the
netic energy in a low Reynolds number, laminar flow with energy resided in the first 50 modes. This was considered a
a limited range of length scales and spatially localised flow relatively poor energy convergence compared to decom-
features may be almost entirely captured with ten modes. positions of other flows where a large proportion of the
At a high Reynolds number, the same flow may be tur- energy was captured in relatively few modes (Delville et al.
bulent, with a variety of length scales present over the 1999; Gordeyev 1999; Hilberg et al. 1994; Kirby et al. 1990;
147
entire domain. In this case, the POD analysis may require Rajaee et al. 1994). The poor energy convergence was
several hundred modes to capture a large contribution attributed to the large variety of scales present throughout
of the kinetic energy in the flow. The analysis also the flow. An appreciable phase correlation was exhibited
de-emphasises infrequent events, although they may be between the time-dependent coefficients of the first two
dynamically important, such as the intermittent phenom- modes, leading the author to conclude that the second POD
ena occurring in turbulent boundary layers, i.e. bursts, mode contained part of the structure described by the first
sweeps and ejections. The temporal averaging often per- mode which had decayed in energy and scale. This
formed during the POD procedure is somewhat arbitrary behaviour was also observed in the spatial structure of the
and not necessarily the best choice for capturing the modes. A significant phase correlation was not observed
dominant behaviour of a system (Graham and Kevrekidis for the time-dependent coefficients of the higher modes.
1996). Finally, since the technique is based on time aver- An investigation of the time-dependent behaviour of the
aged correlations, the POD modes only represent time modes revealed two types of motion present in the recir-
average information about flow structures. culation region of the flow: a swing motion and an
If a random field is homogeneous or periodic in one or expansion and compression motion. These motions
more directions, the eigenfunctions returned from the POD were observed to decrease in amplitude with increasing
analysis are Fourier modes. In such circumstances, Fourier mode order, and were not detectable after ten modes, due
transforms of the data are often used in these directions, but to the limited accuracy of the motion detection technique.
these lead to a poor local description of the flow structures A POD analysis was also performed on the vorticity field
in the physical space. Furthermore, a reduced order model and the results were compared with those obtained from
using Fourier modes must typically retain many modes to the velocity decomposition. The convergence of energy in
adequately capture the dynamics. To overcome this limi- the vorticity eigenvalue spectrum was found to be worse
tation, techniques such as shot-noise decomposition (Arndt than that of the velocity spectrum, with more vorticity
et al. 1997; Herzog 1986; Holmes et al. 1996; Moin and modes required to recover a given energy fraction. How-
Moser 1989) may be applied to gain a physical description ever, this result is misleading, since, in this comparison,
of the flow structure. Alternatively, data preprocessing two different quantities are being compared, i.e. the frac-
techniques known as template fitting and centering (Rowley tion of kinetic energy recovered is being compared to the
and Marsden 2000) avoid imposing Fourier modes as the fraction of enstrophy recovered. Comparisons were also
POD eigenfunctions by effectively performing the POD made between the spatial structure of the vorticity modes
analysis in a travelling reference frame. and the vorticity of the velocity modes. There was good
agreement between the lowest order mode of each, but
1.2 little similarity between the remaining corresponding
Application of POD to wall-bounded flow data modes.
Liu et al. (1994) applied POD to PIV data of a turbulent In another study, Jrgensen (1997) applied POD to the
channel flow to determine whether the eigenfunctions and perturbed and unperturbed flow over a wall-mounted
eigenvalues obtained from a 1-D POD analysis on the fence. PIV was used to acquire 500 instantaneous velocity
outer flow exhibited Reynolds-number-independence fields, each containing 6258 (u, v) components. A 2-D
when scaled by the outer variables. If this was so, then one POD analysis was performed on the entire set of data and
set of eigenfunctions could be used to represent flows over resulted in an 81% and 73% energy recovery for the
a range of Reynolds number. This study was conducted at unperturbed and perturbed cases, respectively, using the
Reynolds numbers of Re=5,378 and 29,935, based on the first 20 modes. This result suggested a greater diversity of
bulk velocity and the channel half width. Their results flow structure from increased 3-D effects as a result of the
indicated that, in the region outside the wall layer, the perturbations.
computed eigenfunctions and eigenvalues exhibited simi- Adrian et al. (2000) used POD as a filtering tool on
larity when scaled by outer variables and, hence, were the PIV data of a fully developed turbulent channel flow at
independent of the Reynolds number. a Reynolds number of Re=5,378 (based on the bulk
Huang (1994) performed a POD analysis on a low velocity and half height of the channel). This technique
aspect ratio backward-facing step (BFS) flow at a Reynolds identified structures based on the size of their spatial or
number of Reh4,300. The PIV data consisted of 250 temporal extent in the flow. The POD procedure was found
instantaneous, sequential velocity fields. Due to the high to be particularly effective at taking into account the
spatial resolution of the data, 691201 (u, v) velocity inhomogeneous spatial distribution of the various eddy
scales, i.e. small eddies close to the wall, large eddies fur- A foreseeable problem arises from using only the vorticity
ther away. A reconstructed/filtered velocity field using the POD modes when a Galerkin projection of Eq. 1 is to be
12 most dominant modes recovered 48% of the total en- performed. The vorticity equation contains both velocity
ergy and 75% of the total Reynolds shear stress, u0 v0 ; in the and vorticity terms, and so, it would be useful to obtain a
flow. relationship between the velocity and vorticity modes. The
A 3-D snapshot POD analysis of a transitional, spatially POD modes in the snapshot technique are based upon
evolving boundary layer was performed by Rempfer and finding the eigenvalues and eigenfunctions of the temporal
Fasel (1994b). Data for this analysis was taken from a correlation matrix, Eq. 13. For velocity data,Cmn is com-
direct numerical simulation by Rist and Fasel (1995). The puted using:
flow domain was split into three regions, each corre- Z
1
148
sponding to approximately two Tollmien-Schlichting Cmnvel um  un dx 2
wavelengths and the corresponding eigenfunctions were M D
determined by applying POD separately to each of the sub- whereas the vorticity formulation requires the calculation
domains. The analysis identified first-order structures that of:
corresponded to structures which are also observed in Z
experiments, namely L vortices. As the flow developed and 1
Cmnvort r  um  r  un dx 3
increased in spatial complexity, the contribution of the M D
first-order structure to the energy of the fluctuating flow
was found to be reduced. The contribution decreased from The presence of products of curl terms suggests that there
95% in the first region, to 86% in the second, to 53% in the is no simple relation between the velocity eigenfunctions
last region. As a consequence, more and more modes were and the vorticity eigenfunctions. It is, therefore, necessary
required to adequately represent the flow field as it to use the Biot-Savart law, shown in Eq. 4, to obtain
evolved, since the energy was now shared more evenly velocity modes from their vorticity counterparts:
between the modes. Z
1 s  xx0
ux  dV x0 s x  x0 4
1.3 4p jsj3
Appropriate choice of variable to use in POD
In the majority of investigations using POD to investigate Here u(x), is the velocity at a point, x, in the flow due to
flow structure, the decomposition has already been per- the contributions of the induced velocity by the vorticity at
formed on velocity data. The aim was to capture most of every other point in the flow, x(x).
the kinetic energy in the flow using the least number of The study by Rempfer and Fasel (1994a) addressed a
modes possible. However, the velocity is not the ideal similar issue when attempting a Galerkin projection of the
quantity to use for coherent structure identification since vorticity equation using their computed velocity modes. In
the deduced structure in the flow will depend on the that study, it was necessary to take the curl of the velocity
velocity of the observer (Perry and Chong 1994). Thus, the modes, which is a somewhat simpler task than using the
modes obtained from a POD analysis of the velocity data Biot-Savart law, particularly in 3-D, to obtain the vorticity
may be misinterpreted as a result of observing them in an modes. These, in conjunction with the existing velocity
inappropriate reference frame. A Galilean invariant modes, were then used as a basis for the projection of the
quantity is preferable for an unambiguous identification of vorticity equation. The matrix r  uk ; r  ul was
coherent structures in the flow. Given the definition of a found to be almost diagonal, indicating that the curl of the
coherent structure by Hussain (1986), the idea that, per- velocity modes did not quite form an orthogonal system,
haps, vorticity may be a better quantity to study than but were almost identical to the vorticity modes.
velocity has been suggested, with the additional benefit Although the vorticity equation has the advantage over
that the vorticity field is a Galilean invariant quantity. A the Navier-Stokes equations in that it has no pressure
POD analysis on vorticity has the potential to be used for term, it requires the knowledge of both the velocity and
the construction of simplified flow models based on vortex vorticity basis functions for a Galerkin projection. If a
skeletons extracted from the modes. Such simplified flow POD analysis on the vorticity was performed, the task of
models have been proposed by Perry and Chong (1987) determining the velocity basis functions would be a sig-
and have aided in determining the type of structures that nificantly more computationally intensive task than
are most important in governing the gross flow behav- extracting the vorticity basis functions from a velocity-
iourthe genetic code of the flow, as put by Perry and based POD analysis.
Chong (1987). However, a quantity suited for coherent structure
The vorticity equation, shown in Eq. 1, rather than the identification, such as vorticity, may give a less efficient
Navier-Stokes equation, could be used to construct and decomposition (in terms of the mean square fluctuating
investigate low-order dynamical models using vorticity value) than a quantity describing flow kinematics, such as
basis functions obtained from POD. The vorticity equation velocity, hence, requiring more modes to completely de-
also has the added advantage of eliminating the pressure scribe the flow. The reasons for applying the POD tech-
term: nique in the first place, i.e. low-order flow descriptions,
coherent structure identification, dynamical flow models,
@x investigation of turbulence quantities etc., will necessitate
u  rx x  ru mr2 x 1
@t the appropriate variable to use.
The idea behind POD analysis is to find a function basis bulent flows (especially wall-bounded flows). It also serves
w(k)(x) that most faithfully represents a random vector to illustrate the effectiveness of vorticity POD analysis as
function, f(x, t), using the form: an objective means of identifying dominant features in
flows.
X
K
k k
f x; t  a t w x 5
k1
2
such that Eq. 5 describes f(x, t) better than any other POD implementation
representation of the same dimension using any other The data used for the POD analysis in the present
function basis. investigation comprises 2-D flow quantities, i.e. (u, v)
Using the definition of the inner product of two velocity components and xz, the out-of-plane vorticity
component, sampled instantaneously in an orthogonal 149
quantities f and w, we obtain:
Z xy coordinate plane over a time period T:. Due to the
f ; w  f x; t w xdx 6 large number spatial points in the domain, roughly
D 8,000, and a greater number of spatial points in the
domain than the number of frames, the snapshot tech-
where * represents the complex conjugate, the energy of
nique (Sirovich 1987) was employed to reduce the
the flow based on the velocity, uand the vorticity x yields
computational effort involved in the POD procedure.
the following results:
The presence of two inhomogeneous directions in the
1 1 flow (streamwise, x and cross stream, y) meant that
Kinetic energy density qu; u qkuk2 7 techniques such as shot-noise decomposition, template
2 2
fitting or centering are not required.
Enstrophy density x; x kxk2 8 Each 2-D field at time tn will be referred to as a snap-
shot, the total number of snapshots being M. The data
Hence, applying POD to u and x will categorise structures exists at discrete points on a rectangular grid with Nx
according to their contribution of kinetic energy and points in the x direction and Ny points in the y direction.
enstrophy in the flow. Note that these quantities are always Note that POD analysis may also be performed on any sub-
positive definite. domain, so that the resulting eigenfunctions are repre-
The data obtained from PIV measurements is well sentative of that sub-domain only. For example, in a
suited to POD analysis. Current experimental PIV tech- turbulent boundary layer flow, POD analysis may be per-
niques are capable of acquiring several hundred instan- formed on the wall and outer layers separately, rather than
taneous two-component, two-dimensional (2C-2D) on the whole flow itself. This approach may be advanta-
velocity fields on which 1-D or 2-D POD may be applied. geous if there is a distribution of scales throughout the
The out-of-plane vorticity field may be obtained from this flow domain. However, the connection between the
data by computing the curl of the velocity field. empirical eigenfunctions obtained from a sub-domain and
The correct identification of a coherent structure in a the entire domain is unclear at this stage and requires
turbulent flow requires the complete 3-D, three-compo- further investigation.
nent velocity and velocity gradient tensor fields (Chong The average snapshot was calculated and subtracted
et al. 1990). Since this information is not available from from each member of the ensemble, leaving only data
PIV data, an alternative means of identifying the presence containing deviations from the mean. Considering the
of coherent structures in the flow is used, based upon the velocity field at this point, u(x,y), the average snapshot is
available information. Throughout this investigation, a calculated using:
large out-of-plane vorticity concentration above the
background vorticity level will be considered as repre- 1X M
senting a coherent feature or structure. x; y
u ux; y; tn 9
M n1
A further advantage of using vorticity in the POD
analysis of a flow is that the overall number of computa- and the resulting adjusted snapshots of data are given by:
tions is reduced, since there is only one component of
vorticity compared to two velocity components (each u0 x; y; tn ux; y; tn  u
x; y 10
vorticity component effectively contains information from
two spatial velocity gradients). Naturally, this only applies In two dimensions, the two-point averaged spatial
to two-component, planar data, such as that obtained from correlation function for the fluctuating quantities is:
PIV measurements. It should be noted that applying POD Z
1 T 0
to vorticity data containing only one component will not Rx; x0 ; y; y0 lim u x; y; t u0 x0 ; y0 ; t dt
T!1 T 0
categorise structures according to their enstrophy content,
but only to their magnitude of out-of-plane vorticity. This 11
quantity is, strictly speaking, a quasi-enstrophy, since only
one component of enstrophy is utilised. However, it will which can be approximated by:
still be referred to as an enstrophy in this investigation. 1X M
The purpose of this study is to investigate the relative Rx; x0 ; y; y0 u0 x; y; tn u0 x0 ; y0 ; tn 12
merits of using velocity or vorticity POD analysis in tur- M n1
since the integration in Eq. 11 is only performed with re- Finally, a reconstruction of any of the original snapshots
spect to t. Following the same procedure as outlined by using an arbitrary number of modes, K, was performed
Sirovich (1987), the correlation matrix for the snaphot using:
technique, C, may be derived as:
ZZ X
K
1 ux; y; tn  u
x; y ank wk x 17
Cmn u0 x; y; tm u0 x0 ; y0 ; tn dxdy 13 k1
M D
The numerical evaluation of the integral in Eq. 13 was
carried out using the trapezoidal rule for each of the 3
integrations, along with an appropriate choice of inner Experimental facilities and diagnostics
150 product for each of the flow quantities. Here, the
common Euclidean inner product or L2-norm was used. 3.1
The eigenfunctions for each mode were determined Water tunnel
using: The experimental investigations were conducted in the
X
M water tunnel shown schematically in Fig. 1a, b. This is a
wk x; y Unk u0 x; y; tn 14 closed-circuit, horizontal facility with five 1-m long
n1 working sections, each of 500500-mm cross-section. Flow
uniformity is achieved with a perforated stainless steel
The expansion coefficients of each mode were calculated plate, a honeycomb and a series of stainless steel screens in
by projecting the original PIV data onto the calculated the settling chamber. The perforated plate (5-mm hole size
modes: with a 23% open area ratio) is placed immediately after the
PN y PN x h 0    i spray system, followed by four screens of decreasing mesh
j1 i1 u xi ; yj ; tn wk xi ; yj
ank PN P h size in the streamwise direction. The honeycomb section
Nx k
   i 15
(12 mm diameter cell size and 120 mm cell width) is in-
y
j1 i1 w xi ; yj wk xi ; yj
serted between the first and second screens to straighten
This expression the flow and remove any mean swirl. A 10:1 contraction
 is simplified
 by making use of the nor-
ratio is used prior to the first working section to further
k l
malisation, w ; w kk dkl : reduce the turbulence intensity by accelerating the mean
Nx h
flow. A maximum flow speed of 775 mm/s is achievable in
Ny
1 XX    i the working sections with the current 53-kW AC motor
ank u0 xi ; yj ; tn wk xi ; yj 16
kk j1 i1 and the in-line centrifugal pump system. The motor pump

Fig. 1. Closed-circuit horizon-


tal water tunnel used in the
experiments. The measure-
ments were taken in working
section3c. (1) Settling cham-
ber. (2) 10:1 contraction ratio.
(3) Working sections. (4) Ple-
num chamber. (5) Rear
observation window. (6) Re-
turn pipeworksuction side.
(7) Filtration isolating valve.
(8) AC motorcentrifugal
pump system. (9) Orifice plate.
(10) Return pipe-
workpressure side. (11)
Water filtration circuit
system speed is controlled with an ABB Sami GS frequency Table 1. Boundary layer properties 18 mm ( 2.25h) upstream
controller, which allows incremental steps in tunnel speed from step. The turbulence intensity values were measured in
the freestream. Water properties were evaluated at 25C and
of 1 mm/s over the entire flow speed range. A plenum 100 kPa for Reh=580 and 35C and 100 kPa for Reh=4,660
chamber attached to the final working section diffuses the
flow and returns it via a 300-mm diameter pipe to the Reh=580 Reh=4,660
pump. Perforated stainless steel plates in the plenum
chamber (10-mm hole size with a 40% open area ratio), U 65 mm/s 425 mm/s
urms
1.1% 1.0%
placed vertically and parallel to the working section walls, U1
Rex 150103 1106
ensure minimal disturbance to the upstream working Red 3,317 24,427
section flow while effectively redirecting the flow through Red* 608 3,177
180 into the return pipe. A diatomaceous earth pool filter Reh 373 2,384
d 45.8 mm 41.9 mm 151
system running in parallel with the main return flow
pipework removes contaminants from the water down to a d* 8.4 mm 5.5 mm
h 5.1 mm 4.1 mm
level of 5 lm. The filtration can be activated at any time, us 3.43 mm/s 18.60 mm/s
but it is never operated during the experiments.

3.2 separation (18 mm 2.25h upstream from the trailing


Backward-facing step (BFS) experiments edge of the plate) for the two BFS flow Reynolds numbers
The flow geometry used in the BFS experiments is shown investigated are shown in Table 1.
in Fig. 2. An 8-mm thick plate positioned 131 mm above
the tunnel floor, ensuring that the turbulent boundary 3.3
layer flow was independent of the tunnel floor boundary Experimental measurement technique
layer, constituted the BFS. This resulted in an aspect ratio Multigrid cross-correlation digital PIV (MCCDPIV) anal-
(step span/step height) and an expansion ratio (y2/y1) of ysis of single-exposed digital images was used to acquire
62 and 1.02, respectively. The experiments were conducted instantaneous, in-plane velocity field measurements of the
at two different freestream velocities,U=65 mm/s and various flows. A Kodak Megaplus XHF camera
U=425 mm/s, yielding Reynolds numbers based on step (1,0001,000 pixels CCD array size, 8-bit greyscale) was
height, Reh=580 and Reh=4,660, respectively. A 5:1 semi- used to acquire singly-exposed images for the low
ellipse was glued on the leading edge of the BFS plate as Reynolds number BFS flow. A maximum framing rate of
shown in Fig. 2 to minimise large-scale flow separation. 30 Hz meant a minimum of 32.9 ms between successive
Given the relatively low Reynolds number of the frames was achievable, hence, its suitability for only the
boundary layer in the experiment (Reg<6,000, see Erm and low speed flow. A PCO Sensicam cross-correlation camera
Joubert 1991), a trip device was used to facilitate its (1,2801,024 pixels CCD array size, 12-bit greyscale) was
transition to the turbulent state. The boundary layer was used to image the high Reynolds number BFS flow. A
tripped using a roughness element that was placed minimum time separation between successive images of
117 mm downstream from the leading edge and consisted 200 ns and an acquisition rate of four frame pairs per
of small acrylic blocks 16-mm long by 5-mm wide by second was achievable with this camera.
6-mm high, glued onto a 2-mm thick polycarbonate strip Each camera was controlled by a separate Pentium
at 14-mm spanwise intervals. The distance from the based PC using image acquisition and timing software.
boundary layer trip to the trailing edge of the plate was The number of images capable of being acquired was
1,850 mm (360h for Reh=580 and 450h for Reh=4,660). dependent on the available PC RAM and the size of frame
The results of Erm and Joubert (1991) show that mean- acquired. The Kodak camera was limited to 32 full-size
flow profiles, broadband-turbulence profiles and spectra image pairs while the PCO camera was limited to 114 full-
are found to be affected very little by the type of trip device size image pairs. Direct acquisition to hard disk permitted
used (wire, distributed grit, pins) for Reg1,000 and above, a larger number of frames to be acquired, albeit at a
indicating an absence of dependence on the flow history reduced framing rate. Synchronisation of the lasers and
for this Reg range. The boundary layer properties prior to digital camera was achieved by sending TTL signals to the

Fig. 2. Flow geometry for the


backward-facing step (BFS).
The boundary layer is
tripped just downstream of
the 5:1 elliptical leading
edge. The step height is
h=8 mm, y1=359 mm,
y2=367 mm. The mean reat-
tachment length is represented
by XR
appropriate devices using the PCs parallel port. The for the low-speed flow where the fluid momentum (and,
timing accuracy of this technique is of the order of 10 ls. hence, the particle momentum) was substantially less.
Illumination of the seeding particles in the flow was The boundary layer properties in Table 1 were obtained
achieved with twin, frequency-doubled Nd:YAG lasers, from the average of 128 instantaneous velocity profiles
each capable of producing a 6-ns, 400-mJ pulse with a using cross-correlation digital PIV. Figure 4 shows the
pulse repetition frequency of 12 Hz. An appropriate mean turbulent boundary layer velocity profile plotted in
combination of cylindrical lenses yielded a 150-mm wide wall coordinates for both the high- and low-speed cases. A
by 2-mm thick collimated light sheet on entry to the test universal turbulent boundary layer profile is also included
section. The flow was seeded with 11-lm hollow glass in each case for comparison. Further details of the up-
spheres having a specific gravity of 1.1. A 105-mm Micro stream boundary layer properties, such as vortical struc-
Nikkor lens used for imaging provided a field of view of ture scales, turbulence intensity profiles and Reynolds
152
approximately 4040 mm ( 5h5h) for the Kodak cam- stress profiles may be found in a PhD thesis by Kostas
era and approximately 5040 mm ( 6.25h5h) for the (2002).
PCO camera. The wall shear velocity, us, was obtained using the
Measurements for the BFS flow were taken at the xy method of Clauser (1956). By plotting the mean velocity
planes shown in Fig. 3, by traversing the light sheet and profile on co-ordinates U/U versus ln(U=y/m) and least
camera across the desired regions. Direct acquisition to squares fitting Eq. 18 to the portion of data exhibiting a
hard disk allowed 1,024 image pairs to be acquired in re- clear semi-logarithmic relationship, us/U may be ob-
gions I and II, and 128 image pairs in the boundary layer tained directly. The universal constants j=0.41 and A=5.0
(bl) region. The time between the image pair acquisitions have been used here:
was set at approximately 2 s to ascertain that the instan-    
taneous velocity fields would be statistically independent. U us U1 y us us us
ln ln A 18
This value was based on a crude estimate of the integral U1 jU1 m jU1 U1 U1
time scale of the turbulent boundary layer. Using a large
The spatial resolution of the imaging system in the high
eddy size d and freestream velocity U, an approximate
Reynolds number case was insufficient to obtain any data
eddy time scale can be calculated;seddy=d/U0.7 s and
in the buffer (5<y+<20) or linear regions (y+<5). A de-
seddy0.1 s for the low and high Reynolds number cases,
viation of the experimental data from the universal profile
respectively. The image pair acquisition rate was also
for small y+ values in both the high and low Reynolds
limited by the hard disk acquisition speed and practical
number cases is most likely due to a bias introduced by a
limitations on the duration of acquisition (17 min for
lack of spatial resolution (i.e. measurement volume too
1,024 frames acquired at 1 Hz). Excessively long acquisi-
large) from the PIV data close to the wall. The spatial
tion times allowed particle sedimentation to develop on
averaging that occurs as a result of the finite interrogation
the plate surfaces and particle drop-out to occur (particles
window size is particularly severe in locations with large
collecting on the screens, honeycomb and settling out in
gradients (i.e. close to the wall) and, in this case, results in
the settling chamber due to the low fluid velocities there
an overestimation of the particle displacements. Further-
and their slight negative buoyancy). This degraded image
more, correlations of the wall surface with itself within the
quality is a result of image flaring at watersurface inter-
interrogation region results in incorrectly determined
faces. The drop-out problem meant that fewer particles
particle displacements.
were present in the flow by the end of the acquisition
The mean flow for both the low and high Reynolds
period compared to the beginning, and was more evident
number BFS flows is shown in Fig. 5. The domain for the
high Reynolds number case is larger owing to the
increased field of view of the PCO camera due to its larger
CCD array. Both velocity fields are an amalgamation of the
mean velocity fields from regions I and II. The mean
reattachment length,XR for both cases differs, with the
high Reynolds number case having the smaller value of the
two. Although not apparent from the presented velocity
field, the high Reynolds number case also contains a sec-
ondary recirculation bubble at the step corner. The low
Reynolds number case has no such secondary recircula-
tion region. The spatial resolution for each statistically
independent velocity measurement is 1.3 mm ( 0.16h) in
both cases.
These single-exposed image pairs were analysed using
the MCCDPIV algorithm described by Soria et al. (1999),
Fig. 3. PIV measurement regions for the backward-facing step which has its origin in an iterative and adaptive cross-
(BFS) flow. The measurement region used for characterising the correlation algorithm introduced by Soria (1994, 1996a,
upstream boundary layer is also shown. Low Reynolds number
BFS flow (Reh=580):xbl=2.5h, ybl=4h, xBFSI xBFSII 5h; 1996b). Details of the performance, precision and experi-
yBFS=2.5h. High Reynolds number BFS flow (Reh=4,680):xbl=4.6h, mental uncertainty of the MCCDPIV algorithm with
ybl=6.25h, xBFSI xBFSII 6:25h; yBFS=2.5h applications to the analysis of single-exposed PIV and
153

Fig. 4a, b. Boundary layer


velocity profile 2.25h upstream
from the step edge for: a the
low Reynolds number
(Reg=373) and b the high Rey-
nolds number (Reg=2,384)
boundary layer flows. The open
circle symbols represent exper-
imental data. Also plotted for
comparison is the universal
boundary layer pro-
file;U+=y+:y+<5;U +=5.0ln
y+)2.9; 5<y+<20 and
U 0:41
1
ln y 5:0 : y > 20

holographic PIV (HPIV) images have been reported in to this, linear particle image distortion (Huang et al 1993)
Soria (1998) and von Ellenrieder et al. (2001), respectively. was used in regions of the flow where there was high shear.
The technique begins with large sampling window sizes The use of these techniques permits a large dynamic range
(SWS) and systematically moves to smaller windows by of velocities to be resolved in the flow whilst preserving
using the displacement estimates from the preceding, spatial resolution by minimising the sampling window
larger window size to offset the small windows. In addition size. The results presented in this paper used a 3232-pixel
SWS as the smallest window size. Vorticity information
was calculated using the technique outlined by Fouras and
Soria (1998) by performing a local, 13-point, 2-Dv2 fit to
the velocity field around the point of interest and then
calculating spatial gradients from this functional fit.
Two errors arise in the calculated vorticity as a result of
this calculation method. A random error, rxz ; exists due to
the error in the velocity field calculation (i.e. the PIV er-
ror) and can be considered to have Gaussian-like prop-
erties. The other is a bias error,xb, which is generally an
underestimation of the peak vorticity, rxmax : Both these
errors depend on the spatial sampling separation of the
data, D, and on the size of a relevant length scale to be
resolved, L, although in an opposing fashion. The com-
plete details of the errors and their determination may be
found in Fouras and Soria (1998) and Kostas (2002). The
vorticity errors for each flow case are summarised in
Table 2. In all cases, independent measurements have been
assumed (i.e. PIV sampling region overlap is disregarded)
Fig. 5a, b. Mean 2-D velocity fields and streamlines for the BFS
flow in the xy plane. a Reh=580. b Reh=4,660. The velocity and the relevant length scale is indicated by L.
measurements have been normalised by the freestream veloc- The entire data set consisted of two-component,
ity,U (only every fourth vector is shown for clarity) instantaneous velocity fields, (u,v), and single-component,
Table 2. Vorticity errors arising from the calculation method.
decomposition data. The technique used to calculate the
The random and bias errors are represented by rxz and xxz b ;vorticity is identical to that mentioned in Sect. 3.3.
respectively. u is the velocity error based on the reference max
The enstrophy spectra for both the velocity and vor-
velocity, Uref ( U and Uo for the BFS flow). L represents a
ticity decompositions for the two flow cases are shown in
characteristic length scale in the flow: dxzsep  mean
 vorticity
q 
Figs. 6 and 7. In all cases, 50 modes were calculated. The
non-dimensional random error transmission ratio 15 D=L
1
spectra show the contribution to the total flow fluctuating
du D xb
enstrophy from each mode. A mode or structure con-
eu Uref % L L rxz keu % xz % tributing a substantial proportion of the total enstrophy to
max

Low Re BFS 0.3 dxzsep 0.32 0.4 6.8 the flow plays a more significant role in the flow dynamics
High Re BFS 1.4 dxzsep 0.64 1.0 27.3 compared to a mode containing less enstrophy. Cumula-
tive sum curves are also included in the plots to illustrate
154
the effectiveness of the decomposition at capturing the
enstrophy in the flow. An efficient decomposition will
instantaneous vorticity fields, xz. Snapshot POD was ap- capture nearly 100% of the fluctuating enstrophy in a
plied to both the velocity and vorticity data for each flow. relatively small number of modes.
The presence of two inhomogeneous directions in the BFS A common trend appearing in the spectra is a more
flow (streamwise, x, and cross stream, y) meant that even distribution of the enstrophy across the modes as the
techniques such as shot-noise decomposition were not flow becomes increasingly more complex. This results in a
required. The phase information of the expansion coeffi- smaller cumulative fluctuating enstrophy sum for a given
cients was obtained directly from the flow field recon-
struction by projecting the original PIV data onto the
calculated modes using Eq. 16.

3.4
POD accuracy and convergence
The number of POD modes accurately resolved in the
analysis were, to a large degree, determined by the number
of data frames available. Also, the presence of random
noise and occasional poor data quality was found to be
reflected in the high-order modes, and essentially repre-
sented background noise. The results of Breuer and Siro-
vich (1991) indicate that for noise levels of the order of
10)2 on a sample size of M=400, the computed eigenvalues
remain accurate for mode numbers at least up to 50, while
the eigenfunctions begin to deviate from the true solution
for mode numbers less than this. Thus, considering that
random noise levels in the velocity and vorticity fields
from this investigation were of the order of 1% and the
sample size was M=1,024, the eigenvalues and eigenfunc-
tions are considered to be accurate up to approximately 50
modes.
Details on the effects of ensemble size, data resolution
and the presence of random noise in the data on the
empirical functions and eigenvalue spectrum calculated
using the POD procedure may be found in the paper by
Breuer and Sirovich (1991).

4
Results and discussion

4.1
POD enstrophy spectra
The term energy relates to different quantities in the
flow when dealing with velocity and vorticity, hence, a
direct comparison between the POD results using velocity
and vorticity cannot be made. A comparison on equal
grounds, though, may be made by taking the curl of the Fig. 6a, b. Mode shape enstrophy distribution for the low
velocity modes to obtain corresponding vorticity modes. Reynolds number BFS flow, Reh=580. a Region I. b Region II. The
solid line is vorticity-based decomposition, the dashed line is
The enstrophy associated with each velocity mode may velocity-based decomposition. The left ordinate axis represents
then be calculated from its corresponding vorticity mode the fluctuating enstrophy of individual modes, the right ordinate
and, now, may be directly compared to the vorticity axis represents the cumulative fluctuating enstrophy sum
Table 3. Enstrophy fractions for the first 25 and 50 modes of the
low Reynolds number (Reh=580) BFS flow for both the velocity
and vorticity decompositions. a Region I. b Region II. eu
represents the cumulative time average fluctuating enstrophy
Modes Velocity Vorticity
hEi hEi hEi hEi
hEfluc i
% hEtot i
% hEfluc i
% hEtot i
%

a 25 68.92 92.61 77.92 94.75


50 79.00 95.01 84.30 96.27
b 25 53.37 75.65 66.58 82.55
50 67.56 83.06 77.18 88.08
155

Table 4. Enstrophy fractions for the first 25 and 50 modes


of the high Reynolds number (Reh=4,680) BFS flow for both
the velocity and vorticity decompositions. a Region I. b Region II.
hEi represents the cumulative time average fluctuating enstrophy
Modes Velocity Vorticity
hEi hEi hEi hEi
hEfluc i
% hEtot i
% hEfluc i
% hEtot i
%

a 25 15.02 52.20 22.37 56.33


50 24.98 57.80 34.79 63.32
b 25 11.08 24.25 16.90 29.21
50 19.04 31.04 27.09 37.89

features (i.e. low-order velocity modes) do not necessarily


have to posses a high enstrophy content. This is evident in
the steppy nature of the individual mode enstrophy
fraction curves.
It is interesting to see that the enstrophy spectra based
on the POD analysis of vorticity and velocity follow a
similar trend in the low Reynolds number flow, but are
Fig. 7 a, b. Mode shape enstrophy distribution for the high significantly different in the high Reynolds number flow.
Reynolds number BFS flow, Reh=4,680. a Region I. b Region II. While POD of vorticity and velocity may both identify
The solid line is vorticity-based decomposition, the dashed line is
velocity-based decomposition. The left ordinate axis represents vortical structures equivalently at low Reynolds numbers,
the fluctuating enstrophy of individual modes. The right ordinate the situation appears different at the higher Reynolds
axis represents the cumulative fluctuating enstrophy sum numbers. As the Reynolds number increases, POD of
vorticity appears to capture vorticity structures more
effectively than an equivalent decomposition of velocity.
number of modes. The cumulative sum of the mode ens- Further studies on high Reynolds number flow cases are
trophy as a fraction of the total enstrophy for each flow for required to investigate this phenomenon.
25 and 50 modes is listed in Tables 3 and 4, respectively. A
decrease in enstrophy content is noticeable between the 4.2
spectra of regions I and II. The increased level of flow Mean flow and POD modes
complexity near the reattachment region may result in a The fraction of energy the average flow contributes to the
less efficient decomposition and would more evenly dis- time averaged total flow energy (hEtot i) is also indicative of
tribute the enstrophy across the modes. The differences in the complexity of the flow. A higher Reynolds number
the fluctuating enstrophy sums are more evident in the (and, generally, a more turbulent flow) will have an
comparison between the low and high Reynolds number appreciable proportion of the energy lying in the fluctu-
cases. This feature has also been demonstrated in the re- ating component compared to a lower Reynolds number
sults of Rempfer and Fasel (1994b) on a transitional, flow (and, hence, a less turbulent flow) where the ratio of
spatially evolving boundary layer. the mean flow energy to total will be greater. The values for
Irregularities in the mode enstrophy distribution curves both the velocity and vorticity of each flow case are listed
for the velocity-based decomposition arise because the in Table 5.
enstrophy associated with the velocity modes is ordered In the low Reynolds number BFS flow, the mean flow
according to kinetic energy content. The corresponding contributes 96% to the time averaged total flow kinetic
enstrophy associated with each velocity mode will not energy in region I and 94% in region II. The high Reynolds
necessarily be ordered in the same way as the kinetic number flow shows that equal fractions of kinetic energy,
energy. This implies that high kinetic energy containing 97%, are contributed to the flow by the mean flow in both
Table 5. Contribution to the time averaged total flow kinetic here). The velocity results seem to contradict the vorticity
energy and time averaged total flow enstrophy from the mean results by having almost equal proportions of kinetic en-
velocity and mean vorticity fields, respectively
ergy associated to the mean flow in both regions.
Kinetic energy Enstrophy The POD modes represent the most common events
hKEmean i hEmean i occurring in the fluctuating velocity/vorticity fields. A
hKEtotal i % hEtotal i %
conceptual aid used in interpreting their physical meaning
BFS low I 96.47 76.23 is to consider them being superimposed upon a base flow
BFS low II 93.59 47.79 (the mean flow here). Modes 1 and 4 of the velocity and
BFS high I 96.83 43.75 vorticity decompositions for both flow cases are shown in
BFS high II 97.11 14.81
Figs. 8, 9, 10 and 11. Having taken the curl of the velocity
modes not only puts them in a convenient reference frame
156
regions. The vorticity decomposition shows a more drastic for flow structure identification, but it also allows a direct
change in mean enstrophy contribution, 76% versus 48% comparison to be made between the velocity and vorticity
for the low Reynolds number flow and 44% versus 15% for modes.
the high Reynolds number flow. These very low values of The vorticity modes tend to highlight vorticity struc-
enstrophy residing in the mean vorticity suggest that re- tures more than the equivalent velocity modes. This is
gion II may contain disorganised vortical motions that particularly evident in the high Reynolds number flows
bear little resemblance to the mean flow (i.e. it is normally where a greater range of scales exists. In the low Reynolds
expected that a base or mean flow exists with fluctuations number BFS flow, where a limited range of scales is
superimposed on it, which does not appear to be the case present, the velocity decomposition does much better at

Fig. 8ad. POD modes in region


I of the low Reynolds number
BFS flow (Reh=580). a Vorticity
mode 1. b Curl of velocity
mode 1. c Vorticity mode 4.
d Curl of velocity mode 4. The
vorticity is normalised by plate
height and freestream velocity,
xz=xzh/U

Fig. 9ad. POD modes in region


II of the low Reynolds number
BFS flow (Reh=580). a Vorticity
mode 1. b Curl of velocity
mode 1. c Vorticity mode 4.
d Curl of velocity mode 4. The
vorticity is normalised by plate
height and freestream velocity,
xz=xzh/U
Fig. 10ad. POD modes in re-
gion I of the high Reynolds
number BFS flow (Reh=4,660).
a Vorticity mode 1. b Curl of 157
velocity mode 1. c Vorticity
mode 4. d Curl of velocity
mode 4. The vorticity is
normalised by plate height
and freestream velocity,
xz=xzh/U

identifying vorticity structures than in the velocity the sign of the reconstruction coefficient effectively con-
decomposition of the high Reynolds number flow. Far trols the arrangement of the positive/negative vorticity
more structure is apparent in the vorticity modes of the bands.
high Reynolds number BFS flow than the corresponding The curl of the velocity modes display similar fea-
velocity modes. tures to the vorticity modes (except for some differences
In all flow cases, the spatial structure of vorticity mode in the sign of the vorticity), but contain substantially
1 is similar to velocity mode 1. However, this is generally less defined bands of vorticity. In addition, the strength
not true for the remaining modes (the differences between of the fluctuating enstrophy associated with the curl of
vorticity mode 4 and velocity mode 4 are representative of the first velocity mode is approximately half that of the
the differences between the other modes) with no one-to- vorticity mode.
one correspondence between the vorticity modes and the The study by Huang (1994) observed similarly shaped
velocity modes. lowest-order modes, both in the velocity and vorticity
A maximum fluctuating enstrophy content of 20% is domain, to those found in region I of this investigation
obtained in region I for the low Reynolds number flow. (see also Kostas 2002). Using a one-term reconstruction of
The modes in each region of the flow are characterised by the first mode, as shown in Eq. 19, they found that the
a large, horizontal band of positive vorticity lying adjacent lowest-order mode was associated with a large-scale, back
to a similarly sized negative, horizontal vorticity band. and forth oscillatory motion of the reattachment points
These bands of vorticity within mode 1 are in close and, hence, a variation in size of the separation bubble:
proximity to the mean separation streamline in region I.
They are closer to the wall in region II. The sign of the  z x; y an1 wx1z x; y
xz x; y; tn 1term x 19
vorticity bands in mode 1 is different for each flow case. A
negative band of vorticity is closer to the wall in the The similarity in the lowest-order modes suggests that a
low Reynolds number case while a positive band exists common motion or flow feature is present in both flows
there in the high Reynolds number case. This has no that contains a significant proportion of the fluctuating
bearing on the flow feature that the mode represents since enstrophy.

Fig. 11ad. POD modes in re-


gion II of the high Reynolds
number BFS flow (Reh=4,660).
a Vorticity mode 1. b Curl of
velocity mode 1. c Vorticity
mode 4. d Curl of velocity
mode 4. The vorticity is
normalised by plate height
and freestream velocity,
xz=xzh/U
4.3
Flow reconstruction accuracy
To investigate whether the 2-D velocity or 1-D vorticity
was the more efficient in the flow decomposition, com-
parisons between the reconstructed vorticity flow fields
and the actual vorticity data were made as a function of the
number of modes used in the reconstruction. In the case of
the velocity decomposition, the curl of the velocity ei-
genfunctions was used as the basis to reconstruct the
vorticity field. Alternatively, taking the curl of the recon-
158 structed velocity field yielded identical results due to the
linear nature of the reconstruction process and the dis-
tributive properties of the curl operator, i.e.:
 
1 2
r  at w1 at w2   
1 2
 at r  w1 at r  w2    20
As a quantitative measure of the accuracy of the POD
reconstruction, the root mean square error over all frames
was calculated between the reconstructed flow field and the
actual data, i.e.:
M ZZ 12
1X 2
exz rec xz x; y; tn  xzrec x; y; tn  dxdy
M n1 D

21
The root mean square reconstruction error, (xz)rec, was
normalised by the maximum absolute value of the vorticity
in the flow, jxzmax j: The results for each flow case are
shown in Figs. 12 and 13.
An interesting feature in the reconstruction error plots
is the initial decrease then increase of the error in region I
of the low Reynolds number BFS flow shown in Fig. 12a.
Here, the vorticity-based reconstruction error rises mar-
ginally above the velocity-based reconstruction error after
approximately 20 modes until about 40 modes, after which Fig. 12 a, b. Reconstruction accuracy for the low Reynolds number
it decreases once again. BFS flow. a Region I. b Region II. The open circles is the vorticity-
based reconstruction, the open squares is the velocity-based
This slight rise then decline of error in the vorticity reconstruction. The root mean square reconstruction
reconstruction of Fig. 12a may be due to poor conver- error,(xz)rec, is normalised by the maximum absolute value of
gence in only some certain modes. Furthermore, since the vorticity in the flow, jxzmax j
the error does not continue to rise by increasing the
number of modes in the reconstruction, it is unlikely
that the entire mode spectrum is affected. However, this 5
can only be valid for the 50 modes that have been Conclusions
computed and the behaviour of the reconstruction error This paper has investigated the merits of using velocity-
may alter when more modes are calculated in the and vorticity-based POD analysis of planar particle image
decomposition. velocimetry (PIV) data to aid in the identification of vor-
The flow reconstruction results indicate that, for all tical structures in a backward-facing step (BFS) flow. The
cases considered here, the vorticity-based flow decom- results suggest that the unambiguous and objective iden-
position produces a reconstructed flow field equivalent tification and description of the dominant vortical struc-
to or better than a flow field reconstructed from a tures in a flow is more effective using a POD analysis of the
velocity-based decomposition using the same number of out-of-plane vorticity field than a POD analysis of the 2-D
modes. The relatively small range of scales and relatively velocity field.
organised flow activity in region I of the low Reynolds Furthermore, the low-order POD modes for the BFS
number flow may be the reason why the reconstruction at the two Reynolds numbers studied (580 and 4,660)
using either basis performs equally well with approxi- bear a common structure, suggesting a large-scale sim-
mately 50 modes in the reconstruction. The energy ilarity independent of Reynolds number. The structure
spectra in Fig. 6a demonstrate that a relatively efficient of the higher-order modes and, hence, smaller
decomposition results when this type of flow is scales were found to depend strongly on the Reynolds
decomposed. number.
Delville J, Ukeiley L, Cordier L, Bonnelt J, Glauser M (1999) Exami-
nation of large-scale structures in a turbulent plane mixing layer.
Part I: proper orthogonal decomposition. J Fluid Mech 391:91122
Erm L, Joubert P (1991) Low-Reynolds-number turbulent boundary
layers. J Fluid Mech 230:144
Fouras A, Soria J (1998) Accuracy of out-of-plane vorticity mea-
surements using in-plane velocity vector field data. Exp Fluids
25:409430
Gordeyev S (1999) Investigation of coherent structure in the similarity
region of the planar turbulent jet using POD and wavelet analysis.
PhD thesis, Department Aerospace and Mechanical Engineering,
University of Notre Dame, Indiana
Graham M, Kevrekidis I (1996) Alternative approaches to the Karh-
unen-Loe`ve decomposition for model reduction and data analysis. 159
Comput Chem Eng 20(5):495506
Herzog S (1986) The large scale structure in the near wall region of a
turbulent pipe flow. PhD thesis, Cornell University, Ithaca, New
York
Hilberg D, Lazik W, Fiedler H (1994) The application of classical POD
and snapshot POD in a turbulent shear layer with periodic
structures. Appl Sci Res 53:283290
Holmes P, Lumley J, Berkooz G (1996) Turbulence, coherent struc-
tures, dynamical systems and Symmetry. Cambridge University
Press, Cambridge, UK
Huang H (1994) Limitations of and improvements to PIV and its
application to a backward-facing step flow. PhD thesis, Technis-
chen Universitat Berlin, Germany
Huang H, Fiedler H, Wang J (1993) Limitation and improvement of
PIV. Part II: particle image distortion, a novel technique. Exp
Fluids 5(34):263273
Hussain F (1986) Coherent structures and turbulence. J Fluid Mech
173:303356
Jrgensen BH (1997) Application of POD to PIV images of flow over a
wall mounted fence. In: Srensen JN, Hopfinger EJ, Aubry N (eds)
Proceedings of the IUTAM symposium on simulation and iden-
tification of organised structures in flows, Lyngby, Denmark, May
1997. Kluwer, The Netherlands, pp 397407
Kevlahan N-R, Hunt J, Vassilicos J (1994) A comparison of different
analytical techniques for identifying structures in turbulence.
Appl Sci Res 53:339355
Kirby M, Boris J, Sirovich L (1990) An eigenfunction analysis of
axisymmetric jet flow. J Comput Phys 90(1):98122
Kostas J (2002) An experimental investigation of the structure of a
turbulent backward facing step flow. PhD thesis, Laboratory for
Fig. 13a, b. Reconstruction accuracy for the high Reynolds Turbulence Research in Aerospace and Combustion, Department
number BFS flow. a Region I. b Region II. The open circles is the of Mechanical Engineering, Monash University Melbourne, Aus-
vorticity-based reconstruction, the open squares is the velocity- tralia
based reconstruction. The root mean square reconstruction error, Liu Z-C, Adrian R, Hanratty T (1994) Reynolds number similarity of
(xz)rec, is normalised by the maximum absolute value of the orthogonal decomposition of the outer layer of turbulent wall
vorticity in the flow, jxzmax j flow. Phys Fluids 6(8):28152819
Lumley J (1967) The structure of inhomogeneous turbulent flows.
In: Yaglam AM, Tatarsky VI (eds) Proceedings of the interna-
tional colloquium on the fine scale structure of the atmosphere
and its influence on radio wave propagation, Moscow, Nauka,
References pp 166178
Adrian R, Christensen K, Liu Z-C (2000) Analysis and interpretation Moin P, Moser R (1989) Characteristic-eddy decomposition of tur-
of instantaneous turbulent velocity fields. Exp Fluids (29):275290 bulence in a channel. J Fluid Mech 200:471509
Arndt R, Long D, Glauser M (1997) The proper orthogonal decom- Perry A, Chong M (1987) A description of eddying motions and flow
position of pressure fluctuations surrounding a turbulent jet. patterns using critical point concepts. Annu Rev Fluid Mech
J Fluid Mech 340:133 19:125155
Berkooz G, Holmes P, Lumley J (1993) The proper orthogonal Perry A, Chong M (1994) Topology of flow patterns in vortex motions
decomposition in the analysis of turbulent flows. Annu Rev Fluid and turbulence. Appl Sci Res 54(34):357374
Mech 25:53975 Rajaee M, Karlsson S, Sirovich L (1994) Low-dimensional description
Bonnet J, Delville J, Glauser M, Antonia R, Bisset D, Cole D, Fiedler H, of free shear flow coherent structures and their dynamical
Garem J, Hilberg D, Jeong J, Kevlahan N, Ukeiley L, Vincendeau E behavior. J Fluid Mech 258:129
(1998) Collaborative testing of eddy structure identification Rempfer D, Fasel H (1994a) Dynamics of three-dimensional coherent
methods in free turbulent shear flows. Exp Fluids 25:197225 structures in a flat-plate boundary layer. J Fluid Mech 275:257283
Breuer K, Sirovich L (1991) The use of the Karhunen-Loe`ve procedure Rempfer D, Fasel H (1994b) Evolution of three-dimensional coherent
for the calculation of linear eigenfunctions. J Comput Phys structures in a flat-plate boundary layer. J Fluid Mech 260:351375
96(2):277296 Rist U, Fasel H (1995) Direct numerical simulation of controlled
Chong M, Perry A, Cantwell B (1990) A general classification of three- transition in a flat-plate boundary layer. J Fluid Mech 298:211248
dimensional flow fields. Phys Fluids 2(5):765777 Rowley C, Marsden J (2000) Reconstruction equations and the
Clauser F (1956) The turbulent boundary layer. Adv Appl Mech Karhunen-Loe`ve expansion for systems with symmetry. Physica D
4:151 119
Sirovich L (1987) Turbulence and the dynamics of coherent Soria J (1998) Multigrid approach to cross-correlation digital PIV and
structures. Part I: coherent structures. Q Appl Math HPIV analysis. In: Thompson MK, Hourigan K (eds) Proceedings
45(3):561571 of the 13th Australasian fluid mechanics conference, Monash
Soria J (1994) Digital cross-correlation particle image velocimetry University, Melbourne, Australia, December 1998
measurements in the near wake of a circular cylinder. In: Pro- Soria J, Cater J, Kostas J (1999) High resolution multigrid cross-
ceedings of the international colloquium on jets, wakes and shear correlation digital PIV measurements of a turbulent starting jet
layers, Melbourne, Australia, April 1994. CSIRO, Australia, pp using half frame image shift film recording. Opt Laser Technol
25.125.8 31:312
Soria J (1996a) An adaptive cross-correlation digital PIV technique von Ellenrieder, K, Kostas J, Soria J (2001) Measurements of a wall-
for unsteady flow investigations. In: Masri AR, Honnery DR (eds) bounded, turbulent, separated flow using HPIV. J Turbulence
Proceedings of the 1st Australian conference on laser diagnostics 2(1):115
in fluid mechanics and combustion, Sydney, Australia, December
1996
160 Soria J (1996b) An investigation of the near wake of a circular cylinder
using a video-based digital cross-correlation particle image ve-
locimetry technique. Exp Therm Fluid Sci 12:221233

Das könnte Ihnen auch gefallen