Sie sind auf Seite 1von 19

Meas. Sci. Technol. 7 (1996) 1498–1516.

Printed in the UK

Coherent structure identification from


the analysis of hot-wire data
P Sullivan† and A Pollard
Department of Mechanical Engineering, Queen’s University, Kingston,
Ontario K7L 3N6, Canada

Received 1 April 1996, in final form 24 June 1996, accepted for publication 4 July
1996

Abstract. This paper describes methods for the identification of coherent


structures from multipoint measurements made in a three-dimensional wall jet
using (a) proper orthogonal decomposition (POD), (b) linear stochastic estimation
(LSE), (c) Gram–Charlier estimation (GCE) and (d) wavelet decomposition (WD).
The comparison of both POD and LSE show that each captures essentially the
same information. The GCE is used to develop a spatially dense mapping of the
instantaneous flow. WD is found to offer a more general method for investigating a
flow field. These methods are applied and the resulting information analysed to
educe structure; it is proposed that the anisotropic growth rate, i.e. faster spreading
in the lateral direction in comparison to the transverse direction, found in the wall
jet is due to large scale structure.

1. Introduction ∆T

Coherent structures can be found from multipoint data


using various interrogation tools: proper orthogonal Hotwire 1
decomposition (POD), linear stochastic estimation (LSE), Vortex moving
Gram–Charlier estimation (GCE) and continuous wavelet at U
transforms (CWTs). In this paper these interrogation tools c
Spatial
are used to evaluate the near field of a three-dimensional Separation
wall jet. This flow is used because it is a combination of ∆s
free shear and near wall flow, both of which have been
found to contain organized motions. Although only one
flow is examined in this paper, this work [1] has been used
to examine data from a channel flow and multipoint cross- Hotwire 2
wire data from a mixing layer. T

v
Averaging Time
1.1. Free shear flows
Coherent vorticity has been used to identify structure in, for u

example, circular jets [2, 3], mixing layers [4], and wakes
[5, 6]. Coherent vorticity is defined here for one spatial Figure 1. Schematic showing technique for finding
dimension and time as, coherent vorticity.
∂v ∂u 1 1v 1u
ωz = − = − (1) large spatial separations between measurement positions in
∂x ∂y Uc 1t 1s
comparison to the temporal separation.
where ωz is the vorticity in the plane transverse to the axis To remedy this, Hussain and Hayakawa [6] suggested
of shear, u and v and x and y are the streamwise and lateral the use of filtered velocity fields, where a short-time average
velocities and directions respectively, Uc is a convection vorticity, ω̃, is defined of the same order as the structure
velocity, and 1t and 1s are separations in time and space, passage time, T ,
see figure 1. The measurement of spatial derivatives, Z 1
1 t+ 2 T
especially with hotwires, is very difficult because of the ω̃(t; T ) = ω(τ ) dτ. (2)
T t− 12 T
† Now at the Department of Mechanical and Industrial Engineering,
University of Toronto, Toronto, Ontario M5S 3G8, Canada. Thus, structure is a region of either positive or negative

0957-0233/96/101498+19$19.50
c 1996 IOP Publishing Ltd
Identification of coherent structures

a1
a2= 2a1

a b
Figure 2. Patterns for (a ) single roller and (b ) double roller. Figure 3. Mexican hat wavelet at scale a1 and scale a2 .

ω̃(t; T ) [7]. Equation (2) can be written discretely as a


spatial and temporal filter of the experimental data
X 
1 T
1v(t, Y ) X T
1t y
ω̃c = − 1u(T , y) (3) x
T t=0
Uc t=0
1s
D/A and Control
z
where T , t are temporal coordinates, and Y , y are spatial Computer Nozzle
coordinates. Therefore, equation (2) is a temporal filter and
Hotwire Rake Rake
also a spatial filter acting over spatial separation, 1s. Bridges (14) 1 2

An improvement to the detection scheme of Hussain


was suggested by Bellin et al [8–10] who introduced
Flow out of page
delocalized conditional sampling (DCS) and applied it to
a mixing layer with a 2:1 velocity ratio. Using rakes of
cross-wires, the measured flow was separated into structure Wall
and non-structure passages, each averaged to determine
its influence on the flow field. After partitioning the
flow, the centre of the structure events could be identified
and phase averages constructed. In order to visualize Figure 4. A general view of the experimental apparatus; x ,
axial direction; y , lateral direction; z , transverse direction.
the measurement, pseudo-flow visualization (PFV) [11]
was used. PFV scales the instantaneously measured flow
field against a colour table, thereby providing an Eulerian new basis pattern, P . The new pattern is then replaced
description of the data, which avoids the need to define a in equation (4) and the operation repeated until minimal
convection velocity. Combining both DCS and PFV, the statistical change occurs in the pattern from iteration to
evolution of a structure in the mixing layer was shown [8]. iteration.
Pattern recognition analysis (PRA) has been used to PRA in its original form modifies the pattern but does
identify regions of coherent velocity. Mumford [12, 13] not identify individual events important to the turbulence
and others [14, 15] proposed the use of the cross-correlation statistics. Thus, there is no clear relationship between
tensor, u(x + 1x, t + τ )u(x, t), as a basis pattern. The the resulting averaged pattern and the original dataset. In
pattern is convolved with a measured flow field the plane jet, Mumford [13] postulated that the identified
structure is responsible for 20% of the mean turbulence
NX X
wires Twindow
energy. In the plane wake, no mention is made of
Ck` = u(i, n + k − 1)P (i + ` − 1, n) (4)
i=1 k=1
the contribution of the resultant pattern to the long time
statistics.
` = 1, . . . , Smax − Nwires + 1 The choice of pattern is also contentious. Single and
double rollers are two fundamental patterns usually chosen
k = 1, . . . , Tmax − Twindow + 1
for plane jets and wakes (figure 2); however, neither pattern
where Ck` is the correlation between the pattern, P , and the is statistically orthogonal. Ferré et al [16] and Mumford
measured velocity field, u; Nwires is the number of wires [12, 13] suggested that it is possible to identify the influence
used in the measurements; Smax is the number of spatial of single and double rollers separately within the same
points for the pattern chosen; Twindow is the total number field. Ferré et al [16] have further postulated that the
of temporal points; n is the temporal length of the pattern problem with separating the influence of the two is a pattern
and Tmax is the number of temporal points for the entire alignment problem. No information is available to support
data set. Every Ck` that exceeds a threshold value of the this supposition.
long time RMS, Crms , is averaged together to construct a Ferré et al [16] examined the influence of the large

1499
P Sullivan and A Pollard

Standard Single Sensorrobe


P Series

Ceramic Tubing 38 mm length Leads 400 mm

Probe Prongs Delrin Block


DUPONT BergCon
15 mm length 25 mm length
Type Connector

LINEAR HO
T-WIRE ARRA
Y
Ceramic Tubing Electrical e
L ads
Aluminum Block
1.6 mm Diameter
400 mm

3.96 mm

Set Screws

Side 7.95 mm
View
End
View

Figure 5. Auspex hotwire rake system.

scale structure on the fine scale turbulence with the fine are an optimal method of flow decomposition, remains
scale turbulence indicator function (FSIF) [17], Fs , to be determined. For certain flows without visually
identifiable structure or low spatial or temporal correlation,
d2 u POD can return basis functions that have no relation to
Fs = . (5)
dt 2 the flow physics. While originally termed characteristic
Fs was calculated for an entire measured dataset and a eddies [19], Lumley [20] later suggested that these
predefined threshold level. Values above the threshold features do not necessarily exist within the flow, merely
were then averaged and compared to the results of PRA; that this technique allows optimal decomposition of a
however, using two different initial patterns with PRA field.
often results in two different final averaged patterns. The
dependence of PRA on the initial conditions make it 1.2. Wall bounded flows
difficult to compare PRA and FSIF.
Delville et al [18] applied the proper orthogonal Willmarth and Lu [21] used quadrant splitting (QS) to
decomposition (POD), described later in section 2.1, to find the statistical influence of the coherent structure
the flow described by Bellin et al [8]. The first mode of in a boundary layer. QS separates the contribution of
the decomposition corresponds to the highest turbulence the instantaneous kinematic Reynolds stress onto four
energy and is taken to be a coherent structure. From quadrants in the u, v hodograph plane. The authors found
the decomposition, it was found that the odd modes, that the largest contribution to the Reynolds stress in the
i.e. first, third, etc are symmetrical about the mixing near wall region occurred in quadrant II. The majority of
layer axis, while the even modes are antisymmetrical the contribution to uv (99%) occurred during only 55%
about this axis. Most of the features along the mixing of the sampling time. QS only identifies the location of
layer axis could be reconstructed with the first mode. individual events within the hodograph plane and does not
However, the passage of a coherent structure from determine whether an event corresponds to either a structure
the low to high speed side of the mixing layer was or a single unrelated peak. Hairpin structures were found
associated with a velocity deficit on the low speed side to be responsible for the ejection of low speed fluid near
and a velocity increase on the high speed side. This the wall [21].
antisymmetrical feature was captured by antisymmetrical Blackwelder and Kaplan [22] introduced variable
modes of POD commencing with the second mode. interval time averaging (VITA), where structure is defined
When the modes identified by Delville et al [18] are as a region of coherent velocity. With VITA, a time
compared with the structures found by Hussain and scale is predefined for the structure passage and used to
Hayakawa [6] in a wake, the even modes seem to calculate a localized RMS. Peaks in the localized RMS
capture vertical slices of the two-dimensional structures, with respect to the long-time RMS of the signal defines
the rolls [6], and the odd modes capture the ribs. structure passage. Since these peaks represent either spatial
Whether the modes correspond to either structure or accelerations or decelerations (if spatial averaging is used)

1500
Identification of coherent structures

Rake 1
Z
Mean Streamwise
Velocity Profile

Outer Region 0.6 h


3.42 h
V Instantaneous view (cartoon)
h of the orifice vortex

Inner Region
X

Side View Spanwise Mean


Velocity Profile
End Wall Z
Re h = 6500 y Wall
x/h = 3
x/h = 5
Orifice x/h = 7

x
Transverse Mean
Velocity Profile
Perspective
Figure 6. Experimental configuration of the wall jet. x /h are axial locations over which whole y –z planes of data were taken.

Figure 7. Decomposition of velocity field: —— measured velocity field, · · · · · · mode 1 of POD, — · — sum of modes 1 and 2
of POD, and – – –sum of modes 1, 2 and 3.

or temporal accelerations or decelerations, VITA will detect stress.


shearing events and neither the peak of the acceleration Sullivan et al modified VITA [24] with the introduction
nor the minima of the deceleration. Shearing events of structure information. Direct comparisons between
do not necessarily have a unique phase relationship with VITA and enhanced VITA (EVITA) applied to an ersatz
structure passage, this makes identification difficult (cf turbulence signal showed that EVITA was less dependent
Robinson [23]). Also, there is subjectivity in defining on either threshold level or averaging time. Applied to
both the structure time scale and the location of the a channel flow [25], EVITA extracted bursting events
peak of the localized RMS. Willmarth and Lu [21] and highlighted various scales involved with bursting.
and Blackwelder and Kaplan [22] both found that Q2 The bursting events were found to be responsible for a
and Q4 events were major contributors to the Reynolds significant portion of the turbulence energy in the field.

1501
P Sullivan and A Pollard

Figure 8. Comparison between the measured velocity field for wire 2 (full curve) and reconstructed velocity for wire 2, using
no information from wire 2 (broken curve).

Figure 9. Comparison between the POD Mode 1 for wire 2 (full curve) and reconstructed velocity for wire 2, with no
information from wire 2 (broken curve).

2. Data analysis is defined as that corresponding to the first eigenmode.


Defining a function, α, the structure φ and the velocity
2.1. Proper orthogonal decomposition (POD) field ui (x, t) can be related:
The proper orthogonal decomposition (POD) is a method α = ui φi . (6)
that decomposes the velocity cross-correlation tensor, Rij ,
into a set of eigenvalues and eigenmodes. The lowest order To maximize the projection of φ(x, t) on the velocity field,
or first eigenmode has the maximum turbulence energy, equation (6) is both squared and averaged,
the second eigenmode has less energy than the first and
more than the remaining and so on. A coherent structure |α|2 = |ui φi |2 (7)

1502
Identification of coherent structures

Figure 10. Axial fluctuation velocity profiles measured with rake 2 at plane 1 (full circles) (x /h = 3) and reconstructed with
0
GCE/LSE (full curve) normalized by the maximum mean centreline velocity (u /Uclmax ). The ordinate is the spatial position
normalized by nozzle height (z /h ). The numbers inside the graphs correspond to lateral positions of the rake: (1)
y /h = −8.63; (2) y /h = −7.20; (3) y /h = −5.77; (4) y /h = −4.34; (5) y /h = −2.91; (6) y /h = −1.48; (7) y /h = −0.05; (8)
y /h = 1.37; (9) y /h = 2.80; (10) y /h = 4.23.

so that |α|2 is representative of the energy in the flow. where the coefficients, an , are,
φi can be found from the integral equation (Lumley ZZ
[19]), an = ui (z, t)φi∗ (z, t) dz dt (12)
ZZ
Rij (z, t, z 0 , τ )φj (z 0 , τ ) dz 0 dτ = |α|2 φi (z, t) (8) and statistically orthogonal [26],
∗ =λ δ .
an a m (13)
n mn
where z 0 is a spatial separation and Rij is the velocity cross-
correlation defined by, The present measurements of the correlation tensor
were done in one spatial dimension and time. The temporal
Rij = ui (z, t)ui (z 0 , τ ). (9) eigenmodes are simply Fourier components. Therefore, the
spatial direction is the only unknown. Fourier transforming
φi can be obtained by recognizing that equation (8) is an equation (8), it is possible to write
eigenvalue problem with the velocity cross-correlation as a Z
kernel. The eigenvalues, λn , are all real. The eigenmodes φij (z, z 0 ; ω)ψj (z; ω) dz 0 = λ(ω)ψi (z; ω) (14)
are orthogonal and satisfy the condition
X where ψj in equation (14) corresponds to φj , in
Rij (z, t, z 0 , τ ) = λn |φin (z, t)|2 . (10)
n equation (8), and φij is the cross-spectrum defined as the
Fourier transform of the cross-correlation,
The eigenmodes are used to represent the flow field as, Z ∞
X φij (z, z 0 ; ω) =
1
Rij (z, z 0 , τ )e−iω(τ ) d(τ ) (15)
ui (z, t) = an φin (z, t). (11) 2π −∞

1503
P Sullivan and A Pollard

3
2
4.0 1
3.4
2.8

z/h
2.3
a 1.7
1.1
0.6

2
4.0
1
3.4
2.8
z/h

2.3
1.7
b 1.1
0.6

4 3 1
4.0 2
3.4
2.8
z/h

2.3
1.7
c 1.1
0.6

time (*3000)

Figure 11. Streamwise turbulence velocity using combined GCE/LSE reconstruction. Spatial density ten times greater than
that measured. Flow is from right to left. Broken curves are negative velocities, full curves are positive velocities. The
reconstructions are for plane 1 (x /h = 3). (a ) y /h = 0.58, (b ) y /h = 2.01 and (c ) y /h = 3.43. Probable structures are noted
with bold numbers on the figure.
and the use of the Fourier transform is used since the flow 2.2. Linear stochastic estimation (LSE)
can be assumed stationary in time. Thus, POD can be used
to identify structure in the z dimension. Linear stochastic estimation (LSE) [27, 28] relates condi-
The decomposed Fourier transformed field can now be tional averages to the correlation functions.
reconstructed The conditional average of a velocity u at a position
X ∗
x + r and time t + τ , when a specified condition occurs at
F (ui (ω; z)) = an (ω)ψni (z; ω) (16) position x and time t, can be written as;
n
û(x, t; r, τ ) = u(x + r; t + τ )|u(x, t). (19)
where F indicates a Fourier transform, and the coefficients,
an (ω), are determined from Normally, conditional measurement at z + z 0 would depend
Z on a trigger condition at z. Stochastic estimation relates,

an (ω) = F(ui (ω; z))ψni (z; ω) dz. (17) without recourse to a trigger condition, the conditional
average to higher powers of the velocity condition u(x)
An inverse Fourier transform is used to obtain the original (reference positions z1 and z2 ), resulting in an estimate of
velocity field, u(z, t; z 0 , τ ),
XZ ∞
ui (z, t) = an (ω)ψni (z; ω) eiωt dω. (18) u(z, t; z 0 , τ ) = A(z 0 , τ )uj (z1 , t) + B(z 0 , τ )uj (z2 , t)
n −∞ +O[u2 (z, t)]. (20)
POD works well in flows where structure is visually The unknown coefficients A, B can be evaluated with either
identifiable. As the flow becomes more random in motion a least squares method or any other appropriate technique
so that less apparent structure exists, the velocity cross- so as to minimize the error between the estimated and actual
correlation, Rij , is of lower magnitude, both spatially velocities at z + z 0
and temporally. As the magnitude of the velocity cross- e = [û − ui (z + z 0 , t + τ )]2. (21)
correlation approaches zero (away from zero temporal and
spatial lag), POD will continue to extract eigenmodes and The condition is met when
eigenvalues. However, these eigenmodes and eigenvalues ∂e ∂e
will have little or no relation to flow structure. = = 0. (22)
∂A ∂B
1504
Identification of coherent structures

3 2
4.0 1
3.4
2.8

z/h
2.3
a 1.7
1.1
0.6

2 1
4.0 3
3.4
2.8
z/h

2.3
1.7
b 1.1
0.6

4.0 2
1
3.4
2.8
z/h

2.3
1.7
c
1.1
0.6

time (*3000)

Figure 12. Streamwise turbulence velocity using combined GCE/LSE reconstruction. Spatial density ten times greater than
that measured. Flow is from right to left. Broken curves are negative velocities, full curves are positive velocities. The
reconstructions are for plane 2 (x /h = 5). (a ) y /h = 0.58, (b ) y /h = 2.01 and (c ) y /h = 3.43. Probable structures are noted
with bold numbers on the figure.

P
Thus, the reconstructed point depends upon unconditional where fi (z 0 ) = λi 8∗i (z 0 )/ j∞=1 λj |8j (z 0 )|2 (Berkooz et al
and not conditional statistics, [29]). Equation (20) (for one spatial reference point) can
therefore be written as
uj (z1 , t)uk (z1 , t)A(z 0 , τ ) + uj (z1 , t)uk (z2 , t)B(z 0 , τ )
= uj (z1 , t)ui (z + z 0 , t + τ ) (23) X

u(z 0 ) = u(z) 8i (z)fi (z 0 ). (27)
uj (z1 , t)uk (z2 , t)A(z 0 , τ ) + uj (z1 , t)uk (z1 , t)B(z 0 , τ ) i=1
= uj (z2 , t)ui (z + z 0 , t + τ ). (24)
Thus, fi is the relative contribution of each mode, 8i , to
From these equations, the coefficients for equation (20) are the conditional average; therefore, as with POD, there is a
found. probability that the eigenmodes will not be associated with
structure, but as random projections on Rij .
2.3. POD and LSE
2.4. Gram–Charlier estimation (GCE)
Using LSE, the relation between the measured and
reconstructed cross-correlations can be written, for one The cross-correlation, Rij (z; r, τ ), for a given spatial
spatially separated reference point and no temporal separation, can be written as,
separation as:
u1 (r, t)u1 (r + 1r, t + τ ) = (P0 + P1 τ + · · · + Pk τ k ) e−Bτ
2

0
u(z)u∗ (z 0 ) = A(z, z )u(z 0 )u∗ (1z). (25) (28)
where k is the order of the expansion and P1...k and B are
From equation (10), it is possible to expand this in terms constants that can be readily found using singular value
of the POD eigenmodes and eigenvalues, decomposition for each velocity pair. This expression
P∞ represents the non-Gaussian terms (large eddies [30]); the
∗ 0 X∞
i=1 λi 8i (z)8i (z )
A(z, z 0 ) = P ∞ 0 )|2
= 8i (z)fi (z 0 ) (26) small scale or Gaussian terms (small eddies [30]) can be
|8
2
i=1 λ i i (z i=1 included by adding A e(−Cτ ) ; however, this term was not

1505
P Sullivan and A Pollard

3 2 1
4.0
3.4
2.8

z/h
2.3
1.7
a
1.1
0.6

4 3 2
1
4.0
3.4
2.8
z/h

2.3
1.7
1.1
b 0.6

4.0 4 2
3 1
3.4
2.8
z/h

2.3
1.7
1.1
c
0.6

time (*3000)

Figure 13. Streamwise turbulence velocity using combined GCE/LSE reconstruction. Spatial density ten times greater than
that measured. Flow is from right to left. Broken curves are negative velocities, full curves are positive velocities. The
reconstructions are for plane 3 (x /h = 7). (a ) y /h = 0.58, (b ) y /h = 2.01 and (c ) y /h = 3.43. Probable structures are noted
with bold numbers on the figure.

4.0
3.4
2.8
z/h

2.3
a 1.7
1.1
0.6

4.0
3.4
2.8
b
z/h

2.3
1.7
1.1
0.6

4.0
3.4
2.8
z/h

c 2.3
1.7
1.1
0.6

time (*3000)

Figure 14. Vorticity contours with the combined GCE/LSE reconstruction, at x /h = 3 and (a ) y /h = 0.58, (b ) y /h = 2.01 and
(c ) y /h = 3.43. Spatial density ten times greater than that measured. Flow is from right to left. Broken curves are negative
and full curves are positive vorticity.

1506
Identification of coherent structures

4.0
3.4
2.8

z/h
2.3
a
1.7
1.1
0.6

4.0
3.4
2.8
z/h
2.3
1.7
1.1
b
0.6

4.0
3.4
2.8
z/h

2.3
1.7
c 1.1
0.6

time (*3000)

Figure 15. Vorticity contours with the combined GCE/LSE reconstruction, at x /h = 5 and (a ) y /h = 0.58, (b ) y /h = 2.01 and
(c ) y /h = 3.43. Spatial density ten times greater than that measured. Flow is from right to left. Broken curves are negative
and full curves are positive vorticity.

included because there was found no computational benefit 2.5. Continuous wavelet transform (CWT)
[1].
In order to compare with other reconstruction The continuous wavelet transform (CWT) can be used to
techniques, it is possible to write equation (28) as, perform a spatio-temporal decomposition of a velocity field
with a set of predefined basis functions. The wavelet, g(t),
X ∞
∂n
Rij = G (29) chosen in this study is the Mexican hat, i.e. the second
n=0
∂τ n derivative of a Gaussian
where G is the Gaussian function and n the order. The
proper orthogonal decomposition (POD) [19] decomposes g(t) = (1 − (t/a)2 ) exp[− 12 (t/a)2 ] (32)
the cross-correlation as
where a is a scale parameter and t is time. This function
X

Rij = λn 8ni 8jn (30) has been shown [1] to be an effective investigative function
n=1 in turbulent flows. By increasing or decreasing the value
where n is the mode of the decomposition, λ an eigenvalue, of a, the basis function in (32) is either stretched or
and 8 the eigenmodes. The eigenmodes and eigenvalues contracted, see figure 3. The Mexican hat wavelet is
are determined from the complete cross-correlation tensor, chosen because it can be used to extract either maxima
Rij . As with POD, the basis functions for this expansion or minima of acceleration within a flow [34]. The
are orthogonal to each other wavelet must meet the compatibility condition [35] of finite
Z ∞ n energy
∂ ∂m Z
n
G m G dτ = δmn . (31) |F (g)|2
−∞ ∂τ ∂τ cg ≡ (2π )n
dω < ∞ (33)
|ω|n
Unlike POD, however, the basis functions used with the
Gram–Charlier expansion are predefined. Using similar where n is the dimension of the space considered
functions [31, 32] there has been found little advantage in and ω is frequency. The wavelet transform is a
the number of polynomials required using either Gaussian convolution between the wavelet function and the velocity
functions or in the number of modes required from POD. field,
It is also interesting to note that unlike POD [33], the ZZ  
basis functions do not degenerate into trigonometric-like t −b
Tg (a, b) = (a n cg )− 2 g∗
1
u(b) dt da (34)
functions for higher modes. a

1507
P Sullivan and A Pollard

4.0
3.4
2.8

z/h
2.3
a 1.7
1.1
0.6

4.0
3.4
2.8
z/h
2.3
1.7
1.1

b 0.6

4.0
3.4
2.8
z/h

2.3
1.7
1.1
c
0.6

time (*3000)

Figure 16. Vorticity contours with the combined GCE/LSE reconstruction, at x /h = 7 and (a ) y /h = 0.58, (b ) y /h = 2.01 and
(c ) y /h = 3.43. Spatial density ten times greater than that measured. Flow is from right to left. Broken curves are negative
and full curves are positive vorticity.

where g ∗ is the complex conjugate of g and (a n cg )− 2 is


1
investigating this flow and applying these interrogation
used to ensure each expansion or contraction of the wavelet tools is to determine the main mechanism(s) behind the
function has the same energy. The inverse transform can anisotropic spread rate in the wall jet, which has identified
be defined as [35], differing ratios of lateral (y) to transverse (z) (for geometry,
ZZ   see figure 4) growth rates, from 3:1 to 8.5:1. This
t −b
u(t) = cg−1 a −(n+1) Tg (a, b)g da db. (35) anisotropic growth rate could be related to flow structure
a that combines characteristics of both free shear and wall
bounded flow. Schwab [37] has found a strong dependence
Wavelet transform data are normally presented as carpet of the flow on the conditions at the jet exit and inferred
plots of the coefficients [34], Tg , which presents data over that effects as far back as the settling chamber could be
its entire range, but it does not necessarily demonstrate important to the subsequent development of the wall jet.
statistical significance of either a scale or range of scales. This suggested the need to examine the flow for features
To find the influence of particular scales, there exists a important to the long time mean that are smeared out by
relationship between the wavelet coefficients and the energy time averaging, which is representative of the type of data
spectrum [36] available for this flow configuration.
Z ∞ Z Z
1 ∞ ∞
|u(t)| dt =
2
|Tg (a, t)|2 dt da
−∞ Z cg 0 −∞ 3.1. Apparatus

= 2E(a) da. (36) The original design and construction of the rig was
0
implemented by Schwab [37]. The supply air is delivered
E(a) can be interpreted as a long time statistical measure by a small fan, mounted on anti-vibration pads; the fan
of the influence of particular scales in the flow, i.e. a local is connected to the settling chamber through a flexible
energy spectrum for each scale. E(a) permits statistically duct. From the settling chamber, the flow exits through
important scales to be determined. a rectangular nozzle. Measurements were made with two
hotwire rakes positioned with a three-dimensional traverse
3. Experiment and data acquired digitally using traditional techniques,
see below. Figure 4 shows the schematic drawing of the
As a vehicle for applying the above techniques, a three- experimental rig and shows the major components. The
dimensional wall jet is used. The main reason for coordinate directions are defined as streamwise (x) (out of

1508
Identification of coherent structures

4.0
3.4
a 2.8

z/h
c b
2.3 a

1.7
1.1
0.6

4.0
3.4
2.8 c b
z/h

a
2.3
b 1.7
1.1
0.6

Figure 17. LSE reconstruction of velocity field for plane 1, position 4, using (a ) second order GCE and (b ) third order GCE.

the page), lateral (y) (parallel to wall) and transverse (z) to ensure statistical independence, one hundred blocks of
(normal to wall). data, each block with 581 samples for each wire, were
Two hotwire rakes, manufactured by Auspex Corpora- collected for a total of 58 100 samples per measurement
tion, were used in the experiment, as shown in figure 5. position sampled at 3 kHz. Each block of data was collected
On each rake, there were seven hotwires, each separated and written to disk before the next block was collected,
by 0.57 h, where h is the exit nozzle height. The total span ensuring statistical independence between blocks of data
covered by each rake is 3.42 h. [38]. Comparison between the mean values and fluctuation
velocities of two independently datasets measured at the
same position with the rakes showed agreement within 1%.
3.2. Data acquisition
Comparing the calibration results with the measured
Two rakes of single-wire probes are used to make data, the error associated with the fit was less than 1%. A
the measurements. Digital sampling is done using calibration was performed at the beginning and the end of
a microcomputer with a Keithley/MetraByte DAS–20 the experiment; the difference between the two calibrations
analogue to digital (A/D) board and simultaneous sample was less than 2%.
and hold. The A/D board was also used to position the The positioning errors with the wires were insignificant.
rakes using stepper motors. Both rakes were moved to the one extreme edge of the
The single wires were calibrated in the potential core of nozzle (y/ h = −8.63 and y/ h = −2.27, for rakes 1 and
the jet at the exit of the calibration rig. A velocity range of 2 respectively). The rakes were then placed against the
2–30 m s−1 was covered. The flow rate was varied over 30 wall, and moved up (+z) 0.1 nozzle heights. Once both
discrete values from which a set of voltage–velocity data rakes were positioned, the computer was then used for all
points were obtained for each wire. The coefficients were remaining movement of the rakes.
evaluated using least squares, for a polynomial fit of fourth The measurements were made with two rakes of seven
order, following [38]. single hotwires. The rakes (figure 6) were positioned
parallel to the wall normal direction or the z axis and
3.3. Error analysis measurements made at three downstream positions, x/ h =
3, 5, 7. Rake 1 was moved in discrete steps in the spanwise
The main thrust of this work is to interrogate the measured direction (along the y axis in figure 6) at y/ h = −2.27,
HW signals for organized flow features. Therefore, in order −0.85, 0.58, 2.01, 3.43, 4.86, 6.29, 7.72, 9.15, 10.57,

1509
P Sullivan and A Pollard

b
a

b
b a

Figure 18. LSE reconstruction of vorticity field for plane 1, position 4, using (a ) second order GCE and (b ) third order GCE.

and Rake 2 was moved discretely in the same direction at linear stochastic estimation (GC/LSE) to remap the data
y/ h = −8.63, −7.2, −5.77, −4.34, −2.91, −1.48, −0.05, onto a denser spatial grid. The Gram–Charlier estimation
1.37, 2.8, 4.23, where y/ h = 0 is the geometric centre has been used here to fit the spatial development of the
plane of the jet. The individual hotwires on rake 1 were cross-correlation tensor [41].
located at z/ h = 0.57, 1.14, 1.70, 2.27, 2.84, 3.40, 3.97 Figure 8 shows the reconstructed velocity field of wire
and the individual hotwires on rake 2 were at z/ h = 0.43, 2, using the statistical information from six wires, excluding
1.0, 1.56, 2.12, 2.69, 3.25, 3.82. wire 2, on rake 1. Of course, any wire could have
been excluded, except the bounding wires 1, 7, 8 and
14. The positive velocity peaks are for the most part
4. Results and discussion well reconstructed, while the negative velocity peaks are
not. Unlike POD, the amplitudes of the stochastically
4.1. Instantaneous velocity profiles
estimated positive valued fluctuation velocities are similar
The data from rake 1, wire 2, which was located at to the measurements. Comparing both mode 1 from the
x/ h = 3, y/ h = 2.01 and z/ h = 1.0, are displayed in POD and the GC/LSE reconstruction (both normalized with
figure 7; also shown is the reconstruction from POD of their respective standard deviations) (figure 9) shows very
this same signal. The first mode of POD clearly captures good agreement. Both reconstructions are essentially in
the peaks in the velocity field and the addition of higher phase, despite some differences in magnitude.
modes only slightly improves the reconstruction. POD
offers a method of mathematically reconstructing a data 4.1.1. Streamwise fluctuation profiles. Figure 10
field; however, unless the first mode contains a significant compares the measurements made with rake 2 and the
percentage of the total energy of the data, a flow structure GC/LSE reconstruction at plane 1 (x/ h = 3). Between
[39, 40] may not be identified. Kevlahan et al [39] suggest y/ h = −8.63 . . . − 2.91, large peaks in u0 occur, which
that large scale structure might correspond to a sum of are of order 0.6 to 1.2, which are significantly larger
modes. While not corresponding to the complete structure, than the measurements. From y/ h = −1.48 . . . 2.80
the peaks reconstructed with POD may correspond to the there is a corrugated profile between the measurement
passage of a large scale structure. Since the magnitude of positions z/ h = 0.43 and z/ h = 2.14. The
the reconstructed signal depends on the spatial resolution corrugated profiles have minima or maxima located at
of the large scale structure, the flow field was decomposed points midway between the reference positions. LSE
with a combination of the Gram–Charlier estimation and reconstructs faithfully the measurement points; however,

1510
Identification of coherent structures

1 2 3

4 5 6

f (Hz)

Figure 19. Wavelet (— · —) and Fourier (——) PSD at x /h = 3, y /h = 0.58. The ordinate is the turbulence energy and the
abscissa is in Hz. (1) z /h = 0.53, (2) z /h = 1.14, (3) z /h = 1.70, (4) z /h = 2.27, (5) z /h = 2.84, (6) z /h = 3.40,
(7) z /h = 3.97. The numbers also correspond to wire number on rake 1.

points in between the reference positions are not strictly unlikely to be related to the flow physics and, thus, these
reconstructed. The reconstructed positions (those between reconstructions are not used later.
the measured positions) are a weighted sum of the Figure 11 shows the projection of a portion of the
instantaneous measurements made at the seven reference measured data set onto a ten times denser spatial grid,
wires (wires 1 . . . 7 or wires 8 . . . 14). For reconstructed showing the temporal development of the turbulence
positions near a reference position, the coefficients (A, velocity field at x/ h = 3 and y/ h = 0.58, 2.01 and 3.43,
B, etc from equation (20)) are near unity for the nearest between z/ h = 0.57 and z/ h = 3.97. All reconstructions
reference wire, and near zero for the remainder of the use the measurements made with rake 1. Note that the
coefficients. Therefore, the corrugated profiles are the entrainment boundary is estimated to be at at z/ h = 2.85
result of low correlation between reference positions. (i.e. between wires 5 and 6) at this downstream location
However, the corrugated profile is also an indicator [37]. Between wire number 1 and wire number 4, the u0
of the structure. If the coincident reference positions velocity contours show flow structures inclined 45◦ to the
were completely uncorrelated, the reconstruction midway wall for all three lateral positions. Examining figure 11(a),
between the reference positions would be zero. As noted three structures are noted (11(a)–1, 11(a)–2, 11(a)–3)
in section 2.3, LSE can be viewed as a weighted sum of which are connected with a bridge of positive velocity fluid,
an infinite number of POD eigenmodes. As POD fails to which because of the location of the velocity cores, suggests
capture in the first mode a significant amount of energy in that structure 11(a)–2 is engulfing structure 11(a)–3. In
this flow, any structure within the flow would be a sum of figure 11(b), there is a bridge of positive fluid between
a number of eigenmodes [39]. Therefore, the corrugated structures 11(b)–1 and 11(b)–2. Since the measurements
profiles are caused by two factors. The reference positions represent only a single component from a single rake, the
are absolutely reconstructed and positions midway between transverse component (z) of the fluid motion can only be
reference locations depend on the correlation between the speculated. Previous work using quadrant analysis [21]
two positions. The magnitude of the reconstructions are has found that structure with positive streamwise velocities

1511
P Sullivan and A Pollard

Plane 1

A
D

Geometric
centreline
of nozzle

B
C
Figure 20. Frequency peaks of the wavelet transform plotted for the entire measurement plane, x /h = 3.

tended to have negative transverse velocities. This suggests Structure 13(a)–1 is engulfing structure 13(a)–2. Structure
that structure 11(b)–1 is in the process of catching up to 13(a)–3 is independent of the other two. There are four
and engulfing structure 11(b)–2. Figure 11(c) shows four positive velocity cores in figure 13(b) (13(b)–1, 13(b)–
potential velocity cores (11(c)–1, 11(c)–1, 11(c)–3, 11(c)– 2, 13(b)–3, 13(b)–4). Structures 13(b)–1 and 13(b)–2
4). It is probable that structure 11(c)–2 and structure 11(c)– are interacting, and structures 13(b)–3 and 13(b)–4 also
3 are the same structure, separated because of the contour seem to be interacting. Figure 13(c) again has four
levels. Structure 11(c)–1 is probably in the initial stages of positive velocity cores (13(c)–1, 13(c)–2, 13(c)–3, 13(c)–
growth and will catch up and engulf structure 11(c)–2, as 4). Structure 13(c)–1 is proceeding downstream, while
seen in figure 11(b). between structures 13(c)–2 and 13(c)–3 there seems to be
Figures 12(a)–(c) are the instantaneous reconstructions some linkage.
in plane 2(x/ h = 5) at the same lateral positions as plane 1. The instantaneous vorticity, defined as ∂u0 /∂z or ωy , is
The transverse spatial extent of the structures is greater than shown in figures 14(a)–(c), for plane 1 (x/ h = 3) at the
that found in the previous plane, extending approximately same lateral positions (y/ h = 0.58, 2.01, 3.43). In all three
to the measurement position of wire 5. Figure 12(a) shows plots there are large ribbons of coherent positive vorticity
a number of positive velocity cores in the flow (the three across the entire temporal range. Moreover, the sequence
with the highest magnitude are labelled 12(a)–1, 12(a)–2, of vortices near the bottom of the figures implies that the
12(a)–3). In figure 12(b), there are again three structures vortex shed from the orifice (see figure 6) maintains some
(12(b)–1, 12(b)–2, 12(b)–3). Structure 12(b)–1 seems to be coherence with ±ω. The mechanism of interaction may
engulfing structure 12(b)–2. Structures 11(b)–3 and 12(b)– be similar to cut and connect [6], which suggests that the
1 are independent of each other. Figure 12(c) has a large fundamental mechanism in this flow is the interaction of the
scale structure (12(c)–1) spanning the temporal extent of the two- and three-dimensional structure that acts to constrain
reconstructions. This is probably two separate structures, growth in the transverse direction. Figures 15 and 16 also
which have already merged. Structure 12(c)–2 is probably show the same ribbons of positive vorticity across the field.
the tail of the original structure being enveloped by 12(c)–1. Note that the contour levels for figure 16 are different than
The instantaneous turbulence velocity reconstructions the contour levels used for figures 14 and 15.
are shown in figures 13(a)–(c). The positive velocity The cross-correlation reconstruction was done with a
cores extend in the transverse directions to approximately second order GCE. Increasing to third order provides very
the measurement position of wire 5. Figure 13(a) shows little difference between the resulting reconstruction and
three positive velocity cores (13(a)–1, 13(a)–2, 13(a)–3). that obtained with the second order GCE (figure 17).

1512
Identification of coherent structures

A
Plane 2
D

Geometric
centreline
of nozzle

B
C
Figure 21. Frequency peaks of the wavelet transform plotted for the entire measurement plane, x /h = 5.

Both figures show similar regions of significant positive shown in the instantaneous velocity reconstruction), both
velocity between wires 4 and 1, with three distinct regions the wavelet and Fourier PSD have significantly higher
(a, b, c) found in both reconstructions. The velocity energy than found for the positions above the location of
contour magnitudes using both the second and third order wire 4. The peak of the wavelet PSD is also shifted to
GC reconstructions are similar and similarly located in both higher frequencies than the peak found with the Fourier
time and space. PSD.
Using the same definition of vorticity as above, the The wavelet decomposition, equation (34) is used to
second and third order GC data were used with LSE to identify scales important within the flow. Combined with
create vorticity maps. These are shown in figure 18. the wavelet PSD, equation (36), scales important to the
The resulting figure shows many differences between long time statistics of the jet are isolated. The wavelet
the two reconstructions, particularly with the two vortex scales have been converted to frequencies [36].
cores labelled a, b. Because of slight differences in the Figure 20 shows a contour plot of the wavelet power
reconstructed velocity field, a completely different vorticity spectral densities (PSD) in the plane normal to the jet axis
field is produced, suggesting two different physical pictures. at x/ h = 3. The measurements use data from both rakes 1
The LSE with the second order GCE results in vortex a and 2. In the figure, four centres of frequency labelled A, B,
leading vortex b. Using third order GCE, a now lags b; C, D are identified. The four peaks in the PSD suggest four
even so, both reconstructions highlight similar regions in cores of quadrant specific motions (QSMs), which implies
both fields. The calculation of derivatives, as noted in that in this region the wall jet contains secondary flows
section 1, is strongly dependent on small changes in the important to the long time statistics.
flow field and thus vorticity is difficult to either measure or Figure 21 shows that the vortex centres, identified in
calculate accurately. figure 20, appear to proceed downstream such that they
remain identifiable at x/ h = 5. The centre of the QSM A
4.2. Fourier and wavelet PSD has been displaced and its frequency significantly increased.
QSMs B, C and D have slightly lower frequencies. The
Figure 19 is the wavelet and Fourier PSD for the high energy motions are still acting to move fluid within
measurements made with rake 1 at x/ h = 3, and y/ h = the y–z plane.
0.58. Both the wavelet PSD and Fourier PSD peak at Figure 22 shows the QSM at x/ h = 7. The energetic
approximately the same location, as expected. Within the frequencies seem to be still associated with lateral and
entrainment boundary (below the position of wire 4, as transverse motions within the flow.

1513
P Sullivan and A Pollard

Plane 3
A
E

Geometric
centreline
of nozzle

B C D

Figure 22. Frequency peaks of the wavelet transform plotted for the entire measurement plane, x /h = 7.

5. Interpretation (as implied by the wavelet PSD). These motions, which


may be representative of streamwise aligned vorticity, must
The data show (a) the existence of a structure in the near necessarily interact with the wall. That fluid is entrained
field of the wall jet (wavelet PSD), and (b) structure, down (negative transverse direction) towards the wall,
similar to the single roller exists within the flow (GC/LSE and out parallel to the wall (parallel with the positive
statistics). Flow visualization made in a facility of the same and negative y axis since symmetry must be maintained)
scale has noted the existence of a ring-like structure within implies that the ring-like vortex issuing from the orifice is
the field [42]. retarded by the wall, which evolves into a structure that
is tilted forward when viewed in a plane parallel with the
streamwise direction, see figure 23. The vorticity must
5.1. Growth rate
diffuse and, thus, the anisotropy in the spreading rate can
The data presented suggest that the anisotropic growth rate only be accomplished by vortex stretching near the wall,
of the jet is determined from the initial structure(s) at the jet which produces a lateral spread in the near wall portion of
exit. Three regions have been defined in the wall jet [43]— the initial orifice vortex.
(a) potential core (PC), (b) shear layer interaction (SLI) and The development of this structure implies that there
(c) final decay (FD). These regions are defined using mean must also be an induced velocity field near the wall.
velocities, i.e. the mean velocity is a constant in the PC Figure 24 shows a cutaway of the identified horseshoe
region (U/Umax ∝ (x/ h)0 ) , decays as a constant power vortex structure with a velocity profile overlaid. The
of downstream distance in the SLI region (U/Umax ∝ vorticity (∂u/∂z) should be dissipated against the wall,
(x/ h)−0.16 ) and is followed by the FD region where the consequently destroying the hairpin, unless the structure
flow develops like a radial jet (U/Umax ∝ (x/ h)−1.1 ). The entrains fluid between the legs of the horseshoe structure
measurements in the present work are limited to the initial and the wall. This would lift the horseshoe structure away
two regions (PC and SLI). Initially, a structure exits from from the wall. This entrained fluid is probably an initiator
the nozzle as a vortex ring/array as diagrammed in figure 6 of near wall structure within the flow.
(x/ h = 3) (that it is either a vortex ring or an array of The initial exit conditions (and the characteristics of the
axially aligned vortices in the four corners of the jet or structure) determine the growth rate of the jet. The structure
even some combination of these, remains debatable [44]). has been found to persist well into the FD region, where
It is believed that the present data suggest that a ring- a 12 Hz passage frequency was found as far downstream
like vortex evolves quickly into quadrant specific motions as x/ h = 49.5 [32]. The dominant structure corresponds

1514
Identification of coherent structures

a
b c downstream. The mechanism of regeneration is similar to
that suggested by Matsuda et al [45], in which the hairpin
Side structure (which, in the present interpretation, is the remains
View of the initial vortex ring generated at the orifice) concen-
trates a region of lateral vorticity near the wall, stretching
x/h = 0
it in the streamwise direction. This ultimately engulfs the
original vortex.

Front View 6. Summary


a

Structure The near field of a three-dimensional wall jet has been


Surface
b
probed using multi point hotwire anemometry. The data
obtained have been supplemented by application of proper
orthogonal decomposition, linear stochastic estimation,
c
Gram–Charlier expansions and wavelet analysis. POD was
used to reveal an underlying structure in the wall jet. With
Figure 23. Development of the initial structure into the final GCE/LSE, the evolution of the structure downstream was
structure, (a ) x /h = 3, (b ) x /h = 5, and (c ) x /h = 7. shown to have an association with the exiting structure from
the rectangular nozzle. There appears, as revealed by the
Fluid entrained power spectral density obtained from wavelet analysis, a
by structure
four-vortex array located in each quadrant of the jet. The
Head
near field evolution of these cores of streamwise vorticity
has been documented. Their link with the orifice generated
Surface of
Horseshoe
vortex is not yet understood.
Vortex

Acknowledgments

This work was supported by the Natural Sciences and


Leg Engineering Research Council through equipment and
operating grants to AP, an equipment grant from Dantec
Corporation and Queen’s University Advisory Research
Council.

Wall
References
Figure 24. Cutaway of the identified structure and the near
wall induced vorticity. Velocity profiles are shown on the [1] Sullivan P E 1995 Coherent structure identification in the
figures. near-field of a three-dimensional wall jet PhD thesis
Queens University, Kingston, Ontario, Canada
[2] Zaman K B M Q and Hussain A K M F 1980 Vortex
to the peak of both the Fourier and wavelet PSD. It has pairing in a circular jet under controlled excitation Part
a significant amount of turbulence energy and is important I. General jet response J. Fluid Mech. 101 449–91 Part 3
[3] Hussain A K M F and Zaman K B M Q 1980 Vortex
for the redistribution of fluid.
pairing in a circular jet under controlled excitation Part
The saddle-backed velocity profiles found in this study II. Coherent structure dynamics J. Fluid Mech. 101
and in the studies of Schwab [37] and others are due to the 493–544 Part 3
influence of this initial structure. The vortices act to move [4] Hussain A K M F and Clark A 1981 On the coherent
fluid to the lateral edges of the jet from the transverse edge structure of the axisymetric mixing layer. A flow
visualization study J. Fluid Mech. 104 263–94
of the jet. As the structure transforms into the large hairpin-
[5] Hayakawa M and Hussain A K M F 1989
like structure, the saddle-backed velocity profiles start to Three-dimensionality of organized structure in a plane
disappear. Schwab found that the saddle-backed velocity turbulent wake J. Fluid Mech. 206 375–404
profiles existed through x/ h = 10, after which they started [6] Hussain A K M F and Hayakawa M 1987 Eduction of
to decay and eventually disappeared by x/ h = 30. From large-scale organized structures in a plane turbulent
the initial stages and development of the structure exiting wake. J. Fluid Mech. 180 193–230
[7] Hussain A K M F 1981 Role of coherent structures in
the nozzle, it is postulated that the structure transforms into turbulent shear flows Proc. Indian Acad. Sci. Eng. Sci. 4
a symmetric (about the transverse axis) structure, much like pt 2 129–81 Invited lecture given at Asian Congress of
the one postulated by Matsuda et al [45]. Fluid Mechanics (Bangalore, India)
The instantaneous GC/LSE reconstruction show inter- [8] Bellin S, Delville J, Vincendeau E, Garem H and Bonnet
action between structures, which is characteristic of a cut J P 1993 Large scale structure characterization in a 2d
mixing layer by pseudo-flow visulization and delocalized
and connect mechanism for the regeneration of structure. conditional sampling Eddy Structure Identification in
One flow structure has been found to engulf another and Free Turbulent Shear Flows ed J Bonnet and M Glauser
the resulting structure is suggested to continue its evolution (Dordrecht: Kluwer) pp 91–102

1515
P Sullivan and A Pollard

[9] Bellin S 1988 Contribution à l’étude des structures dans la [28] Adrian R 1988 Linking correlations and structure:
zone turbulente développée d’une couche de mélange stochastic estimation and conditional averaging Near
plane, par développement de techniques d’analyse des Wall Turbulence ed S Kline and N Afghan (Washington,
données issues de peigne de fils chauds Diplome DC: Hemisphere) pp 420–36
D’Etudes Approfondies Centre d’Etudes [29] Berkooz G, Elezgaray J, Holmes P, Lumley J and Poje A
Aérodynamiques et Thermiques, Poitiers, France 1992 The proper orthogonal decomposition, wavelets
[10] Bellin S 1991 Etude experimentale des structures and modal approaches in the dynamics of coherent
coherentes d’une couche de mélange plane turbulente de structures Technical Report FDA-92-12 Cornell
fluide incompressible PhD Thesis Centre d’Etudes University
Aérodynamiques et Thermiques, Université de Poitiers, [30] Payne F 1969 Generalized Gram-Charlier method for curve
France fitting statistical data AIAA J. 7 2045–46
[11] Tabatabai M, Kawall J G and Keffer J F 1987 Flow [31] Berkooz G, Holmes P and Lumley J 1993 The proper
visualisation using hotwire anemometry DANTEC Info. 4 orthogonal decomposition in the analysis of turbulent
[12] Mumford J C 1982 The structure of the large eddies in flows Annual Review of Fluid Mechanics ed J L Lumley,
fully developed turbulent shear flow. Part I. The plane M Van Dyke and H L Reed (Palo Alto, CA: Annual
jet J. Fluid Mech. 118 241–68 Reviews Inc.)
[13] Mumford J C 1983 The structure of the large eddies in [32] Sullivan P and Pollard A 1996 Wavelet decomposition in a
fully developed turbulent shear flow. Part II. The plane three-dimensional wall jet and an enclosed channel. In
wake. J. Fluid Mech. 137 447–56 preparation
[14] Kopp G, Kawall J and Keffer J 1993 On the uniform [33] Sung H and Adrian R 1994 Karhunen–Loève expansion of
distortion of plane turbulent porous-body wakes Ninth the derivative of an inhomogeneous process Phys.
Symp. on Turbulent Shear Flows (Kyoto, Japan) August Fluids. 6 2233–5
1993 [34] Farge M 1992 Wavelet transforms and their applications to
[15] Ferré J A and Giralt F 1989 Pattern recognition analysis of turbulence Annual Review of Fluid Mech. ed J L
the velocity field in plane turbulent wakes J. Fluid Lumley, M Van Dyke and H L Reed (Palo Alto, CA:
Mech. 198 27–64 Annual Reviews Inc.)
[16] Ferré J, Mumford J, Savill A and Giralt E 1990 [35] Gill A G 1992 The wavelet transforms as an analysis tool
Three-dimensional large-eddy motions and fine-scale for structure identification in molecular clouds PhD
activity in a plane turbulent wake J. Fluid Mech. 210 Thesis Queen’s University, Kingston, Ontario, Canada
371–414 [36] Higuchi H, Lewalle J and Crane P 1994 On the structure of
[17] Hedley T and Keffer J 1974 Turbulent/non-turbulent a two-dimensional wake behind a pair of flat plates
decisions in an intermittent flow J. Fluid Mech 64 Phys. Fluids 6 297–305
625–44 [37] Schwab R R 1986 An experimental and numerical
[18] Delville J, Bellin S and Bonnet J P 1990 Use of proper investigation of three-dimensional jet flows from
orthogonal decomposition in a plane turbulent mixing sharp-edged orifices PhD Thesis Queen’s University,
layer Turbulence and Coherent Structures ed M Lesieur Kingston, Ontario, Canada
and O Métais (Dordrecht: Kluwer) [38] George W K 1978 Processing of random signals
[19] Lumley J 1967 The structure of inhomogeneous turbulent Proc. Dynamic Flow Conf. (Skovlunde, Denmark,
flows Atmospheric Turbulence and Radio Wave 1978) pp 757–800. Appendices by Beuther J and
Prpogation ed A Yaglom and V Tatarsky (Moscow: Lumley J
Nauka) pp 166–76 [39] Kevlahan N, Hunt J and Vassilicos J 1993 A comparison of
[20] Lumley J 1970 Stochastic Tools in Turbulence (New York: different analytical techniques for identifying structure in
Academic) turbulence Eddy Structure Identification in Free
[21] Willmarth W and Lu S 1972 Structure of the Reynolds Turbulent Shear Flows ed J Bonnet and M Glauser
stresses near the wall J. Fluid Mech. 55 (Dordrecht: Kluwer) pp 311–24
[22] Blackwelder R and Kaplan R 1976 On the wall structure of [40] Sullivan P and Pollard A 1992 Preliminary investigation of
the turbulent boundary layer J. Fluid Mech. 76 89–112 coherent structures in a three-dimensional wall jet using
Part I PFV and POD Proc. IUTAM Symp. Eddy Strucutre
[23] Robinson S 1990 Kinematics of turbulent boundary layer Identification in Free Turbulent Shear Flows (Poitiers,
structure PhD Thesis Stanford University France) October 1992
[24] Sullivan P, Day M and Pollard A 1994 Enhanced VITA [41] Sullivan P, Pollard A, Delville J and Bonnet J P 1996
techniques for turbulent structure identification Exp. Two-dimensional Gram-Charlier reconstruction of
Fluids 18 10–16 velocity correlations Phys. Fluids at press
[25] Liu Z, Landreth C, Adrian R and Hanratty T 1991 High [42] Goddard P 1995 flow visualisation of a rectangular jet
resolution measurement of turbulent structure in a Fourth Year Undergraduate Thesis Queen’s University,
channel with particle image velocimetry Exp. Fluids 10 Kingston, Ontario, Canada
301–12 [43] Sforza P M and Herbst G 1970 A Study of
[26] Glauser M N and George W K 1992 Application of three-dimensional, incompressible turbulent wall jets
multipoint measurements for flaw characterization Exp. AIAA J. 8 276–82
Thermal. Fluid Sci. 5 617–32 [44] Pollard A and Schwab R 1993 The field of a rectangular
[27] Adrian R 1977 On the role of conditional averages in free jet Exp. Fluids submitted for publication
turbulent theory Turbulence in Liquids (Princeton, NJ: [45] Matsuda H, Iida S and Hayakawa M 1990 Coherent
Science Press) structures in a three-dimensional wall jet J. Fluids Eng.
ASME Trans. 112 462–7

1516

Das könnte Ihnen auch gefallen