ARTICLE IN PRESS

JOURNAL OF
SOUND AND
VIBRATION
Journal of Sound and Vibration 300 (2007) 50–70
www.elsevier.com/locate/jsvi

Axial flow fan broad-band noise and prediction

Thomas Carolus, Marc Schneider, Hauke Reese
Department of Fluid and Thermodynamics, University of Siegen, 57068 Siegen, Germany
Received 7 November 2005; accepted 19 July 2006
Available online 10 October 2006

Abstract

Two prediction methods for broad-band noise of low-pressure axial fans are investigated. Emphasis is put on the
interaction noise due to ingested turbulence. The numerical large eddy simulation (LES) is applied to predict the unsteady
blade forces due to grid generated highly turbulent inflow; the blade forces are then fed into an analytical two-dimensional
acoustic ducted source model. A simple semi-empirical noise prediction model (SEM) is utilized for indicative comparison.
Finally, to obtain a database for detailed verification, the turbulence statistics for a variety of different inflow
configurations are determined experimentally using hot wire anemometry and a correlation analysis. In the limits of the
necessary assumptions the SEM predicts the noise spectra and the overall sound power surprisingly well without any
further tuning of parameters; the influence of the fan operating point and the nature of the inflow is obtained. Naturally,
the predicted spectra appear unrealistically ‘‘smooth’’, since the empirical input data are averaged and modeled in the
frequency domain. By way of contrast the LES yields the fluctuating forces on the blades in the time domain. Details of
the source characteristics and their origin are obtained rather clearly. The predicted effects of the ingested turbulence on
the fluctuating blade forces and the fan noise compare favorably with experiments. However, the choice of the numerical
grid size determines the maximal resolvable frequency and the computational cost. As contrasted with the SEM, the cost
for the LES-based method are immense.
r 2006 Elsevier Ltd. All rights reserved.

1. Problem statement

The aeroacoustic noise of low Mach number axial fans either results from the encounter of blades with
space and/or time-varying disturbances (interaction noise) or the flow over the blades themselves (self-noise).
Axial-flow fans, used e.g. in air conditioning systems, are often subjected to poor inflow conditions which
range from steady state but spatially asymmetric velocity profiles—due to imperfect intake geometry—to time-
varying ingested vortices, turbulence or secondary inflow distortions. The resulting periodic and/or random
forces cause tonal and/or broad-band interaction noise. On the other hand, if the inflow is homogeneous in
time and space, the force fluctuations due to the turbulent boundary layer (TBL) on the blade surfaces and
their interaction with the trailing edge (TE) and the vortex shedding from the blunt TE cause the unavoidable
self-noise of the fan which in most cases is broad-band. This study is intended to contribute to the prediction
of broad-band fan noise with an emphasis on the interaction noise due to ingested turbulence.

Corresponding author. Tel.: +49 271 740 2386; fax: +49 271 740 2388.
E-mail address: [email protected] (T. Carolus).

0022-460X/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jsv.2006.07.025
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 51

Nomenclature V_ volume flow rate

w flow velocity in the rotating frame of
a bar width of square-mesh array reference
AC correlation area x, y, z duct coordinate system (stationary frame
c0 speed of sound of reference)
c flow velocity in the stationary frame of z number of blades
reference g blade stagger angle
cx,m averaged axial flow through duct without d boundary layer displacement thickness
obstructions, (V_ =ðpr2a  pr2i Þ) L integral length scale
C chord length x1, x2, x3 blade coordinate system (rotating frame
D impeller diameter, inner diameter of duct of reference)
( ¼ 2ra) r fluid density
f frequency j flow rate coefficient
F blade force
Ma circumferential Mach number ( ¼ u/c0) Subscripts
n impeller speed
OAPWL overall sound power level u circumferential
PSDcx power spectral density of axial flow x axial ( ¼ in the x-direction)
velocity fluctuation
PSDF power spectral density of blade force Superscripts
fluctuation
0
PSDsp power spectral density of blade surface velocity fluctuations
—
pressure fluctuation spatial (circumferential) average
PSDW power spectral density of the acoustic
power Abbreviations
PSDL level of PSD
ReC Reynolds number based on chord length BLR boundary layer removal
and w1 IT incident turbulence
SrLx Strouhal number based on Lx and cx LES large eddy simulation
Srd Strouhal number based on d and w1 HC honeycomb
Tu turbulence intensity NI natural inflow
ra radius of impeller ( ¼ D/2) RPG1, RPG2 fine, coarse square-mesh array
ri radius of hub SEM semi-empirical noise prediction model
t time TBL turbulent boundary layer
u circumferential velocity of impeller at ra TE trailing edge
(tip velocity) TCS turbulence control screen

Many semi-empirical noise prediction models (SEM) have been published for fan or wind turbine noise
prediction. A simple early model is due to Sharland [1] which was fundamental for many subsequent studies
on fan noise. Sharland started from a flow containing rigid surface. The surface radiates sound into the free
field due to wall pressure fluctuations from incident turbulence (IT), TBL/TE interaction and TBL/blade
surface interaction. Among other authors, Költzsch extended Sharland’s model integrating more empirical
data and taking into account a duct as proposed by Morfey [2], see Ref. [3]. The crucial assumption is always
acoustic compactness, i.e. the characteristic dimensions of the surface are much smaller than the acoustic
wavelength. Other semi-empirical models for a variety of airfoil self-noise mechanisms, for instance based on
the experimental work on airfoil noise radiation by Brooks et al. [4], have been used by Lowson [5–7] and
more recently by Moriarty and Migliore [8].
The compact source assumption is removed by non-compact linearized theories. Amiet [9] and Paterson and
Amiet [10] developed analytical models for airfoil response to IT and noise which provide absolute level
 ARTICLE IN PRESS
52 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

prediction without recourse to empirical or adjustable constants. The noise due to turbulent flow past a TE
was treated by Ffowks-Williams and Hall [11] and Amiet [12]. The blade response function for pressure fields
convected past the TEs of a linear cascade of blades was investigated by Glegg [13], including a duct by Glegg
and Jochault [14]. A further extension of Amiet’s original formulation of TE noise is given by Roger et al. [15],
accounting for all the effects of a finite chord length and aiming at the prediction from wall pressure
fluctuations computed by incompressible large eddy simulation (LES).
Due to the rapid increase of computational power the numerical prediction of unsteady flow phenomena
from first principles is becoming more and more feasible, yet, in complex flow fields such as in
turbomachinery, it is still a demanding task. Current schemes focus more on the prediction of tonal noise
due to blade/wake interaction. Algermissen et al. [16] used the unsteady Reynolds average navier–Stokes
(URANS) method to predict the noise radiation of automotive cooling fans. Struts had been placed directly
upstream of the fan impeller. Kouidri et al. [17] investigated the response of swept blades to a gust. Their
numerical simulation yields steady and unsteady loading on the blades; the Ffowcs Williams & Hawkings
formulation was used to derive the tonal part of the acoustic spectra of the unducted impeller. However, in the
case of—on average—spatially uniform but temporal unsteady inflow the URANS-method fails, and the
prediction of broad-band noise sources requires a high level numerical method. Thus, You et al. [18] used LES
to investigate the tip clearance flow. Even more complex simulations such as the unsteady flow field in a
complete mixed flow pump have been carried out by Kato et al. [19].
One goal of this study is the use of the numerical LES for broad-band fan noise prediction. For further
indicative comparison a simple SEM is chosen. (As compared to the expensive LES, semi-empirical and
analytical models are far more practical in an industrial environment, yet, to the knowledge of the authors,
still not widely in use.) A disadvantage of any semi-empirical model based on the acoustic analogy is the
unavoidable need of input: A knowledge of the statistics of the turbulent flow around or approaching the
blades is required. Since the noise due to IT is very often dominant in industrial fan applications, it is a second
goal of this study to provide measured turbulent inflow parameters for a variety of inflow configurations such
as extremely clean inflow—achieved by a hemispherical turbulent control screen, honeycombs (HCs) and a
duct boundary layer suction flow—, natural inflow (NI) and several levels of highly turbulent inflow generated
by a grid type ‘‘turbulence generator’’.

2. Experimental setup

2.1. Investigated fan impeller

The investigated fan consists of a typical low-pressure axial impeller with six NACA 4509-blades (Fig. 1),
which is running in a duct-type housing without guide vanes. The impeller outer diameter is D ¼ 299 mm. For
all experiments the speed is n ¼ 3000 rpm (corresponding to a tip speed Mach number Ma ¼ u=c0 ¼ 0:14, with
the circumferential speed of the impeller u ¼ pDn and the speed of sound c0). The Reynolds number ReC,
based on chord length C and mean flow velocity in the rotating impeller (roughly w1, see Fig. 3), varies from
92,000 at the hub to 170,000 at the blade tip. To indicate the fan operating point the dimensionless flow
rate coefficient
V_
j¼ , (1)
uðp=4ÞD2
in preference to volume flow rate V_ is used. The operating point of maximum efficiency (design operating
point) corresponds to j ¼ 0:179. This operating point or its corresponding volume flow rate was selected for
most measurements within this study. Further details of the fan are listed in Table 1.

2.2. Test stand

The impeller is tested in an anechoically terminated duct test stand (following ISO 5136 [20]) with a duct
diameter D ¼ 300 mm (Fig. 2). The suction side of the test stand is surrounded by an anechoic chamber, not
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 53

Fig. 1. Impeller investigated (if not stated otherwise, dimensions in mm).

Table 1
Fan: Data at design point and important geometrical parameters

Speed Z 3000 rpm

Volume flow rate V_ 0.59 m3/s
Total to static pressure rise Dpts 170 Pa
Total to static efficiency Zts 45%
Impeller diameter D ¼ 2ra 0.299 m
Number of blades z 6
Hub/tip ratio ri/ra 0.45
Radial tip clearance s/D 0.0017
Trailing edge thickness d/D 0.0017
d/C 0.0083
Fluid density (air) r 1.2 kg/m3

shown there. The volume flow rate is controlled by an adjustable throttle at the downstream side of the duct
and measured by a calibrated hot film sensor.
The impeller is mounted downstream of a variety of optional devices, by which the inflow turbulence is
controlled (Fig. 2 and Table 2). The NI through a bellmouth-type intake nozzle is the basic configuration; the
ingested turbulence is not controlled by any further means. The porous wall duct section with an external
suction flow is designed to remove the duct boundary layer (BLR) and thus to reduce wall generated
turbulence. Grids are used in fluid flow studies either for the production or reduction of turbulence, and either
for the creation or elimination of large-scale velocity or pressure non-uniformities. A first measure to reduce
the turbulence of the NI is a layer of HC placed upstream of the impeller. Moreover, a turbulence
 ARTICLE IN PRESS
54 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

Fig. 2. Test stand (schematically); hemispherical turbulence control screen (TCS), turbulence control devices (HC, RPG1, RPG2) and
duct boundary layer removal (BLR) optional; the downstream microphone with nose cone and slit tube, anechoic termination and
adjustable throttle are not shown; D in mm.

hemispherical control screen (TCS) surrounding the intake nozzle, with a double layer of honeycomb and fine
wire mesh according to Refs. [21,22], is used to generate an inflow with extremely low turbulence. Finally, two
square-mesh arrays of square bars (RPG1 and RPG2) are designed to produce a highly turbulent inflow.
Combinations of the devices allow for a large variety of inflow conditions.
The turbulence parameters are measured in a reference plane close to the leading edge plane of the impeller;
the fan blades are removed but the hub is still present and non-rotating. An auxiliary fan far downstream
provides the through-flow. Turbulence measurements are taken using one or two single hot-wire probes (5 mm
diameter, 2 mm sensing length). They are oriented normal to the main flow direction and coupled to a
standard multi-channel constant-temperature anemometry system (Dantec CTA StreamLine). The sampling
rate is 20 kHz. In the case, where two probes are used, the signals are captured synchronously. The recorded
data have not been corrected to account for the finite length of the hot wire.
The sound power radiated into the free field on the suction side of the fan is derived from the sound pressure
measured at several positions on a control surface around the inflow [23]. The duct sound power is measured
by a microphone downstream of the impeller. In order to compensate for pseudo-sound and duct mode
effects, the microphone is equipped with a slit tube and a nose cone (as recommended in Ref. [11]). The overall
sound power of the fan is the sum of both. Spectra are measured with a resolution of Df ¼ 3.125 Hz. The
amplitudes of sound power spectra are presented in terms of their power spectral density (PSD) or its level
PSDL. The reference pressure p0 ¼ 2  105 Pa, the reference sound power P0 ¼ 1012 W and the reference
band width Df0 ¼ 1 Hz.

2.3. Flow kinematics

Any local flow velocity is decomposed in a time-mean c and a fluctuating part c0 ðtÞ. Since all quantities are
relevant in the stationary as well as in the rotating system, the different coordinate systems and flow velocity
components have to be distinguished carefully (Fig. 3). The x, y, z-coordinate system is stationary and has its
origin on the center line of the duct. The x1, x2, x3-system is fixed to the rotating blade and has its origin at
midchord and midspan. Given that the single hot-wire probe in the stationary system provides information on
the velocity fluctuations in x-( ¼ axial) direction c0x , the projection w02 ¼ c0x cos g is the relevant quantity,
acting as the source of fluctuating loads on the blades according to the thin-airfoil linearized theory. As it is
the case for any low-pressure axial fan, the blade stagger angle g is rather small (at the outer blade sections less
than 201). Thus w02 is not far from c0x and—as long as the fan operates close to its design point—w1 not far
from the circumferential speed of the impeller u.
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 55

Table 2
Investigated inflow configurations

No. Designation Description Duct section length

A (D)
(D ¼ 300 mm)

1 NI Natural inflow through 1.02

bellmouth intake
2a BLR Boundary layer removal 3.02
2b by suction flow through 1.02
a porous wall section
3 HC Plane layer of 30 mm 3.02
honeycomb in front of
the impeller

4 RPG1 Fine square-mesh array 3.02

of square bars upstream
of the impeller
a  a ¼ 10  10 mm,
b  b ¼ 65  65 mm
5 RPG2 Coarse square-mesh 3.02
array of square bars
upstream of the impeller
a  a ¼ 15  15 mm,
b  b ¼ 60  60 mm
6 TCS Hemispherical 1.02
turbulence control
screen in front of the
nozzle, see Fig. 2

7 TCS+BLR Combination of 6 and 2 1.02

8 TCS+HC Combination of 6 and 3 1.02
9 TCS+HC+BLR Combination of 6 and 3 and 2 1.02

2.4. Turbulent inflow quantities

The local turbulence intensity in the blade leading edge plane

qffiffiffiffiffiffiffi
c0x 2
Tu ¼ (2)
cx
 ARTICLE IN PRESS
56 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

Fig. 3. Coordinate systems in the stationary and rotating frame of reference with decomposition of flow velocities (velocity triangle) and
axial and circumferential correlation length of an ingested eddy.

qffiffiffiffiffiffiffi
is measured with a single hot wire probe, c0x 2 being the rms value of the axial velocity fluctuations. The axial
correlation length is determined assuming Taylor’s ‘‘frozen turbulence’’-hypothesis which allows to set
q=qt ¼ cx q=qx, Hinze [24]. The integral of the auto-correlation function
c0x ðtÞc0x ðt  tÞ
Rxx ðtÞ ¼ , (3)
c0x 2
yields the integral time scale
Z 1
I¼ Rxx ðtÞ dt, (4)
0

which is multiplied by the local axial mean velocity cx to obtain the axial integral length scale
Lx ¼ Icx . (5)
Because of the small stagger angle g of the fan blades the length scale of a turbulent eddy in spanwise
direction x1 (Fig. 3) of a blade corresponds more or less to the circumferential integral length scale Lu in the
stationary system. This length scale is determined from the correlation of two hot wire signals. For any radius
r, one probe is kept fixed at the 3 o’clock position (corresponding to y ¼ 01), while the second is moved step by
step counterclockwise. Thus the circumferential distance ry is increasing from 0 mm to a maximum value,
given here at 541 through practical consideration. The maximum of the correlation coefficient function rxy (t)
is determined for each data pair (using the Matlabs Vers. 6.1 routine xcorr) and used to calculate the
circumferential length scale
Z 1
Lu ¼ maxðrxy ðtÞÞr dy. (6)
0

‘‘Infinity’’ is replaced by the maximum circumferential distance [25]. Although in general the length scales are
a function of the frequency, the integral length scales according to Eqs. (5) and (6) are an estimate of the
largest turbulent structure encountered in the flow.

3. Semi-empirical noise prediction model (SEM)

The SEM for ducted turbomachinery used here is based on Morfey [2], Költzsch [3] and Sharland [1] and
summarized as follows: A uniform hard-walled annular duct is assumed to contain a uniform steady flow on
which the acoustic excitation is superimposed. Reflections at the open ends are neglected, as if the duct were
infinite. Fundamental is the spectral density of the acoustic power W radiated from z uncorrelated broad-band
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 57

sources as given in [3, p. 355, Eq. (10)]:

dW ðf Þ p z f
PSDW ðf Þ  ¼ C PSDF ðf Þ, (7)
df 4 rc0 ra ð1  ðri =ra Þ2 Þ2
2

ri is the hub radius, ra ¼ D=2, and r is the fluid density (air). The radiation factor C accounts for mean flow in
both the +x- and x-direction, but it is set to 1 here because of the low axial flow Mach number. The sources
are represented by the force fluctuation on a blade, PSDF, which can be expressed in terms of the fluctuating
surface pressure distribution PSDsp and their correlation area Ac
ZZ
PSDF ðf Þ ¼ PSDsp ðf ; x1 ; x3 Þ Ac ðf ; x1 ; x3 Þ dx1 dx3 . (8)
A

In this paper the integral in Eq. (8) is approximated by a summation over several strip-type blade elements
from hub to tip (typically nine) which are assumed to radiate incoherently. A further assumption in this model
is the acoustic compactness of the incoherently radiating source regions. That means that the chord length and
spanwise extension of a blade strip should be clearly smaller than the wavelength of the radiated sound; this
can be a rather constraining prerequisite.
For the two noise mechanisms ‘‘IT’’ and ‘‘TBL’’ the surface pressure fluctuations and the correlation
area are now estimated using semi-empirical sub-models. A different empirical sub-model is used for the
‘‘TE’’.

3.1. IT

As depicted in Fig. 3 a turbulent eddy approaches the blade with the steady state velocity w1 in the
rotating blade coordinate system. For the time being, w1 is assumed constant as the eddy passes the blade. Due
to the turbulence the eddy has a superimposed velocity fluctuation perpendicular to the surface
w02 ðx1 ; x3 ; tÞ5w1 . Following [1], the phase-delay between changes in angle of attack a0 and the induced
changes in lift force is neglected. (This quasi-steady approximation may provide overestimates.) Then w02
causes a local fluctuating lift force which corresponds to a local pressure difference between the two sides of
the blade
w21
Dp0 ðx1 ; x3 ; tÞ ¼ c0L ðx1 ; x3 ; tÞr
, (9)
2
where the lift coefficient c0L can be written in terms of a0 ðtÞ due to w02 ðtÞ by defining a lift curve slope, F, such
that
c0L ðtÞ ¼ c0L ða0 ðtÞÞ ¼ F arctanðw02 ðtÞ=w1 Þ  Fðw02 ðtÞ=w1 Þ. (10)
Hence, the mean squared pressure difference becomes
 2 2  2 2
w w0 2 w
Dp0 2 ¼ c0L 2 r 1 ¼ F2 22 r 1 . (11)
2 w1 2
As in Ref. [1] we set F ¼ 0:9p. Defining PSDw2  dw02 2 =df the desired power spectral density of the pressure
fluctuations due to oncoming turbulence becomes

dDp0 2 1
PSDsp  ¼ ð0:9pÞ2 r2 w21 PSDw2 . (12)
df 4
For small blade angles g PSDw2 can be approximated by the so-called one-dimensional energy spectrum
PSDcx  dc0x 2 =df of the velocity fluctuation c0x .
Numerous empirical correlations are published which give the non-dimensionalized spectrum
 
PSDcx
PSD Lcx  10 log dB, (13)
cx Tu2 Lx
 ARTICLE IN PRESS
58 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

as a function of the Strouhal number

f Lx
SrLx ¼ . (14)
cx
The two correlations which have been integrated in the model described here are:

(i) A polynomial curve fit to an extensive series of measurements by Költzsch in Ref. [3]:

X
4
PSD Lcx ðSrLx Þ ¼ ak ðlogðSrLx ÞÞk1 dB (15)
k¼1

with a1 ¼ 9.784, a2 ¼ 19.001, a3 ¼ 5.548 and a4 ¼ 0.060;

(ii) a formula from Hinze [24]
4
PSD Lcx ðSrLx Þ ¼ 10 log dB. (16)
1 þ ð2pSrLx Þ2

Thus, given any values for Tu and Lx , the power spectral density of the velocity fluctuations is known as a
function of frequency. Eqs. (13) and (14) gave rise to experimentally evaluate the statistical quantities Tu and
Lx of the turbulent inflow to the fan impeller.
The correlation area is assumed to be in the order of the size of an oncoming turbulent eddy, AC / L2 . The
frequency associated with a ‘‘frozen’’ turbulent eddy of length L moving with w is f ¼ w=L, thus,
AC / ðw1 =f Þ2 . Sharland [1] eventually proposed
w21
AC ¼ . (17)
ð2pf Þ2

3.2. TBL

The wall pressure fluctuations, generated within the TBL, are estimated by a polynomial curve fit to
measured wall pressure fluctuations on a flat blade [26] and adjusted by an empirical constant to adapt the
data to rotating turbomachinery blades according to Költzsch in [3]:
!
2 3  0:01
PSDsp ðf Þ ¼ r w1 d . (18)
1 þ 4:1985 Srd þ 0:454 Sr6d

The relevant Strouhal-number is now defined with the boundary layer displacement thickness d as
Srd ¼ f d =w1 . (19)
The displacement thickness is estimated from the well-known flat plate approximation
d ¼ C0:0518 Re0:2
C . (20)
The correlation area is adopted from Mugridge [27]:
8
>
> 1 1 2
>
> w1 C ðra  ri Þ for pSrC p2;
>
> 5p f
>
>
< 2
21 15
AC ð f Þ ¼ 5p2 w1 2 Cðra  ri Þ for 2ppSrC p p ; (21)
>
> f
>
>
>
> 6 1 15
> 3
: p4 w1 f 3 ðra  ri Þ
> for ppSrC :
p

Here the Strouhal-number is defined with the chord length C as

SrC ¼ fC=w1 . (22)
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 59

3.3. TE

Out of numerous models (see Section 1) the simplest approach for the TE has been chosen. Based on the
extensive measurements by Brooks et al. [4] on NACA 0012 profiles Lowson [5] developed an empirical
correlation of the 1/3-octave sound pressure due to the turbulent TE flow in a distance r from the TE
1
p0 2ac;TE ðf Þ ¼ 7:079  1012 p20 d Ma5 ðra  ri Þ Gðf Þ (23)
r2
with the spectrum shape function
4ðf =f peak Þ2:5
Gðf Þ ¼ , (24)
ð1 þ ðf =f peak Þ2:5 Þ2
and the peak frequency
0:02w1 Ma0;6
f peak ¼ . (25)
d
p0 ¼ 2  105 Pa is a reference pressure. In Eq. (23) an additional directivity function which takes account for
the dipol radiation pattern has been omitted. Thus, an upper limit of the sound power is obtained by
integration of the sound power as if the sound is radiated from the TE as a spherical wave
4p
PSDTE ðf Þ ¼ 7:079  1012 P0 d Ma5 ðra  ri Þ Gðf Þ, (26)
Df 1=3oct
where P0 is the reference sound power 1012 W and Df 1=3oct is the bandwidth of each 1/3-octave band. To
calculate the sound power of the complete impeller, Eq. (26) is multiplied by the number of blades z.

4. Large-eddy flow field simulation (LES)

The numerical LES method is exclusively applied to the inflow configuration RPG2 (no. 5 in Table 2). Its
restricted usage for highly turbulent inflows stems from the fact that any self-noise mechanism would require
an extremely fine resolution of the boundary layers in the machine which was beyond the available computer
capacity. The LES-code named Frontflow Blue was developed by Kato and has been successfully applied (see
e.g. Ref. [28]). It is based on a finite element discretization of the flow equations. The restriction to
incompressible fluids is not relevant here because of the low value of the (circumferential) Mach number
Ma ¼ 0:14.

4.1. Governing equations and numerical scheme

The governing equations are the spatially filtered continuity equation and the Navier–Stokes equations in
Cartesian coordinates. The velocity fluctuations are decomposed in the numerical grid scaled velocity, resolved
by the LES (subscript ‘‘gs’’), and the sub-grid scale velocity (subscript ‘‘sgs’’) which is still modeled by a
turbulence model: c0 ðtÞ ¼ c0gs ðtÞ þ c0sgs . In this study, the sub-grid scale velocity fluctuations are modeled with a
dynamic Smagorinsky model. Since the motion of large eddies is resolved directly, a high accuracy and
stability of the discretizing method is required. Conventional schemes with a high numerical dissipation are
not appropriate because they will damp out the motion of eddies. Therefore a streamline upwind finite element
formulation is used to discretize the governing equations. This scheme is of second-order accuracy in terms of
both space and time. Details are given by Kato et al. [28,29].

4.2. Numerical grid and boundary conditions

The numerical grid is divided into four sections (Fig. 4). The first section of the grid covers the inlet section
which is a cylindrical duct. At the inlet, the boundary condition is set to a uniform axial velocity. Because of
the complex geometry of the square-mesh array RPG2, a static oversetting is used between the mesh array and
 ARTICLE IN PRESS
60 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

Fig. 4. Numerical grid of the complete flow domain; only every third grid point and mesh line is plotted for clarity.

the inlet section. The oversetting is dynamic between the stationary mesh array and the impeller with its
rotational frame of reference. The numerical grid in a blade channel is divided into five blocks. For the near
blade region, an O-topology is used. In order to reduce the computational cost the tip clearance has been
neglected. The hub of the impeller is extended down to the outlet. At the outlet boundary surface the static
pressure is set to zero. To avoid backflow over the outlet region during the iteration process the geometry of
the outlet duct is modified such that the flow is partially accelerated. This is achieved by a diffuser-type
expansion of the cross-sectional area with a subsequent contraction to the original size of the cylindrical duct-
shaped housing. As compared to the impeller grid the outlet grid is coarser. This requires a third oversetting
region. Non-slip wall conditions are applied to the remaining boundaries of the flow domain.
As initial conditions the velocity and the pressure is set to zero in the whole flow field. To stabilize the
simulation at the beginning and to avoid pressure shock waves, an exponential start-up function is used which
increases the velocity at the inlet boundary step by step. The time increment DtSolv, which is mainly determined
by the stability of the simulation, is set such that 10,000 increments correspond to one single revolution of the
impeller, i.e. DtSolv ¼ 2  106 s for n ¼ 3000 rpm.
In order to ensure an acceptable CPU time the overall number of hexahedral elements for the whole flow
domain is limited to approximately 5 millions. Still, on a Hitachi SR 8000 computer with 8 nodes and 8 CPUs
each, the simulation of one impeller revolution required approximately 40 h. Further details can be found
in Ref. [30].

5. Results

5.1. Statistics of the inflow turbulence

For four typical inflow configurations Fig. 5 depicts fringe plots of the measured local time-mean axial
velocity cx in the reference plane. cx is non-dimensionalized with the averaged axial velocity without any
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 61

1.0

0.9

0.8

0.7
y/ra [-]

0.6

0.5

0.4

0.3
(a) (b)
0.2
cx /cx,m
1.5
1.0
1.4
0.9 1.3

0.8 1.2

0.7 1.1
y/ra [-]

1.0
0.6
0.9
0.5
0.8
0.4
0.7
0.3
0.6
(c) (d)
0.2 0.5
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
z/ra [-] z/ra [-]

Fig. 5. Measured distribution of the local time-mean axial velocity downstream of (a) RPG2, (b) NI, (c) TCS, (d) TCS+HC+BLR.

obstructions cx;m ¼ V_ =ðpr2a  pr2i Þ. RPG2 produces a distribution which is extremely non-uniform due to the
wake/vortex structure downstream of the square-mesh array (Fig. 5a). The NI yields a more uniform
distribution (Fig. 5b). When the turbulence control screen TCS is used the core is very uniform but the
boundary layers are clearly detectable at the duct and hub wall (Fig. 5c). TCS in combination with a HC and
BLR yield a uniform distribution with slightly lower velocities at the duct wall, where the suction is applied
(Fig. 5d). Each plot shows distortions of the maps at the right-hand side boundary at the 2 o’clock position,
corresponding to y ¼ 01. These are caused by the wake of the second fixed hot-wire probe, placed slightly
upstream at y ¼ 01 as well. Therefore, for the following quantitative analysis of the data the region from
y ¼ 0–51 for all radii is excluded. Fig. 6 throws further light on details of the flow field in the reference plane. It
presents—as a function of the radius—circumferentially averaged (indicated by the bar above the symbol)
time-mean velocities, turbulence intensities and length scales. Clearly, the velocity is almost independent of all
turbulence control devices and not significantly influenced by the radial position. However, turbulence
intensity and integral length scales vary considerably. (Note, that the boundary layer could not be fully
detected because of the absolute size of the hot wire probe.) RPG2 produces turbulence intensities as high as
21% (Fig. 6a). TCS reduces Tu down to less than 1% in the core region but not in the vicinity of the duct wall
(Fig. 6c). The combination of TCS, HC and BLR yields an extremely uniform Tu-distribution (Fig. 6d). At the
same time this combination produces the smallest integral length scales. Clearly, the length scales are the
 ARTICLE IN PRESS
62 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

24 48 24 48

20 40 20 40

Λx [mm], Λu [mm]
Λx [mm], Λu [mm]
16 32 16 32
cx [m/s], Tu [%]

cx [m/s], Tu [%]
12 24 12 24

8 16 8 16

4 8 4 8

0 0 0 0
0.5 0.6 0.7 0.8 0.9 1.0 0.5 0.6 0.7 0.8 0.9 1.0
(a) r/ra [-] (b) r/ra [-]

24 48 24 48

20 40 20 40
Λx [mm], Λu [mm]

Λx [mm], Λu [mm]
16 32 16 32
cx [m/s], Tu [%]

cx [m/s], Tu [%]

12 24 12 24

8 16 8 16

4 8 4 8

0 0 0 0
0.5 0.6 0.7 0.8 0.9 1.0 0.5 0.6 0.7 0.8 0.9 1.0
(c) r/ra [-] (d) r/ra [-]

Fig. 6. Measured turbulent inflow parameters as a function of the radius in the reference plane; (a) RPG2, (b) NI, (c) TCS,
(d) TCS+HC+BLR; J: cx , &: Tu, n: Lx , ,: Lu ; the bars indicate a spatial averaging of the quantity in circumferential direction.

largest with NI as well as with TCS. As shown in Section 5.2 and as expected, the combination of TCS, HC
and BLR produces the lowest fan noise. Data for all turbulence control devices investigated within this project
are published in Ref. [31]. Table 3 lists numerical values of the circumferentially averaged turbulent
parameters and the ratio Lx =Lu from tip to hub. In general, Lx and Lu have the same order of magnitude, but
in many cases Lx tends to be slightly larger than Lu . According to Roach [32] several previous workers have
found Lx p1:9L
ffiffiffiffiffiffiffiffiffiffiffiu . In case of mesh generated turbulence, Roach also presented an empirical correlation
Lx =a ¼ 0:2 Dx=a, where Dx is the distance downstream from the turbulence control device and a is the bar
width of the mesh array. Measurements made agree reasonably well with Roach’s results.
Three one-dimensional energy spectra for the inflow configurations TCS, BLR and RPG2, non-
dimensionalized as defined in Eq. (13), are presented in Fig. 7. They were measured with a single hot wire
at a monitoring point not too far away from the duct wall. The spectra coalesce and are correctly predicted by
either Eq. (15) or (16). Although not shown, the data for all other inflow configurations and monitoring points
fit well.
A comparison of the LES-predicted and measured distribution of the time-mean axial velocity in the
reference plane is shown in Fig. 8. The wake/vortex structure due to the coarse square-mesh array RPG2
agrees quantitatively and qualitatively well with the experimental data. Further LES-predicted results are
presented at three different monitoring points in the reference plane (Fig. 9). The three monitoring points are
selected as indicated in Table 2: P1 just behind a bar of RPG2, P2 behind a joint of two bars, and P3 behind an
open area between two bars. At these points the local turbulence intensity, the axial integral length scale and
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 63

Table 3
Measured turbulence statistics

Inflow configuration Range of the measured turbulent parameters from hub to tip
(spatially averaged in circumferential direction; flow rate
corresponding to j ¼ 0.179)

No. Designation Description Duct section Lx (mm) Lu (mm) Lx =Lu () Tu (%)
length A (D)

1 NI Natural inflow through 1.02 12–35 5–13 1.2–3.3 0.6–14.2

bellmouth
2a BLR Boundary layer removal 3.02 9–28 5–13 1.2–2.6 0.6–9.4
2b 1.02 7–28 4–18 1.1–3.4 1.0–11.6
3 HC Plane layer of honey- 3.02 3–15 4–12 0.7–1.8 1.4–13.5
comb
4 RPG1 Fine square-mesh array 3.02 9–16 5–16 0.9–2.2 12.5–20.9
a  a ¼ 10  10 mm,
b  b ¼ 65  65 mm
5 RPG2 Coarse square-mesh 3.02 10–18 7–14 1.1–1.8 16.7–21.3
array
a  a ¼ 15  15 mm,
b  b ¼ 60  60 mm
6 TCS Hemispherical 1.02 7–28 3–40 0.2–2.0 0.4–8.0
turbulence control
screen
7 TCS+BLR Combination of 6 and 1.02 1–8 3–24 0.1–3.3 0.5–5.0
2b
8 TCS+HC Combination of 6 and 3 1.02 2–7 2–14 0.3–2.1 1.3–13.4
9 TCS+HC+BLR Combination of 6 and 3 1.02 2–7 2–8 0.4–1.2 1.4–3.3
and 2b

20

10

0
PSDL*cx [dB]

-10

-20

-30

-40

-50
0.01 0.05 0.1 0.5 1 5 10 50 100
SrΛx [-]

Fig. 7. Collapse of the non-dimensionalised energy spectra of the axial velocity fluctuations in the reference plane at r=ra ¼ 0:96, y ¼ 01;
measurement with TCS, BLR and RPG2 (___________), correlation from Hinze Eq. (15) (- - - - - - - - -), correlation from Költzsch Eq. (16)
( ).

the time-mean axial velocity are calculated. Given the uncertainties in the measurements, the agreement
between prediction and measurement is very satisfactory.
The power spectral density of the axial velocity fluctuations at monitoring point P2 is depicted in Fig. 10. As
compared to the experimental data the LES results agree well up to the ‘‘cut-off frequency’’ due to the filtering
of the LES-method, corresponding to a Strouhal number of 0.4 (E300 Hz).
 ARTICLE IN PRESS
64 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

cx /cx,m
1.5
1.0
1.4
0.9 1.3

0.8 1.2

0.7 1.1
y/ra [-]

1.0
0.6
0.9
0.5
0.8
0.4
0.7
0.3 0.6
(a) (b)
0.2 0.5
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
z/ra [-] z/ra [-]

Fig. 8. Local mean velocity distribution downstream of RPG2 in the reference plane; (a) measurement, (b) LES-prediction.

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0.0
P1 P2 P3 P1 P2 P3 P1 P2 P3
Tu [-] Λx /a [-] cx /cx,m [-]

Fig. 9. Local turbulence intensity, integral length scale and time-mean flow velocity downstream of RPG2 in the reference plane at three
monitoring points P1, P2 and P3 (see Table 2): LES-prediction (white bars) vs. measurements (black bars).

5.2. SEM-predicted and measured fan noise

A merit of the SEM is that the spectral contribution from each source to the overall spectrum can be
quantified (Fig. 11). The TE contributes the least, the TBL is the dominant source in the high-frequency range
and for turbulent inflow—depending on its turbulence intensity—in the low-frequency range. The influence of
Tu is considerable (Fig. 12). Even for moderate turbulence intensities, the turbulent inflow mechanism
becomes dominant in the low-frequency range.
Figs. 13 and 16 show measured and SEM-predicted sound power spectra. Strictly speaking—because of the
assumption of acoustically compact sources—the comparison of predicted and measured broad-band noise is
valid only in the frequency region up to 1 kHz, where the chord length is clearly smaller than the wavelength of
the radiated sound. Given all other assumptions (duct without reflecting boundaries (Eq. (7)), quasi-steady lift
force response to ingested turbulence (Eq. (10)), Költzsch’s empirical constant in the wall pressure fluctuations
beneath a TBL (Eq. (18)) and the sound power from a simplified TE sound field (Eq. (26))), the SEM predicts
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 65

20

10

PSDL*cx [dB]
-10

-20

-30

-40

-50
0.01 0.05 0.1 0.5 1 5 10
SrΛx [-]

Fig. 10. Non-dimensionalized power spectral density of the axial velocity fluctuations in the reference plane at monitoring point P3;
measured (___________), LES-predicted ( ).

70

60

50 IT IT+TBL+TE
PSDLW [dB]

40 TBL

30 TE

20

10

0
100 500 1000 5000 10000
f [Hz]

Fig. 11. Contributions to the overall spectrum from the three noise mechanisms TE ( ¼ trailing edge), TBL ( ¼ turbulent boundary layer)
and IT ( ¼ incident turbulence) as predicted with SEM; data correspond to inflow configuration NI; ref 1012 W/Hz.

80

70

60
Tu = 20 %
PSDLW [dB]

50

40 TBL+TE
Tu = 2 %
30

20 Tu = 8 %
10

0
100 500 1000 5000 10000
f [Hz]

Fig. 12. Influence of turbulence intensity on the overall spectrum as predicted with SEM; L is set to a constant value of 35 mm;
ref 1012 W/Hz.
 ARTICLE IN PRESS
66 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

the spectra reasonably well without otherwise tuning any parameters. It is worth mentioning that BLR reduces
the broad-band levels between 1 and 5 kHz. As seen in Fig. 16, RPG2 increases the low frequency broad-band
levels up to 1 kHz.
Surprisingly all measured spectra show peaks at multiples of BPF which of course are not predicted by the
SEM. In order to eliminate sources of these tonal components in the experiments a number of measures have
been investigated such as an increased precision of the geometry of the inflow nozzle and the tip clearance, a
minimization of the wobble and the unbalance of the impeller and a variation of the impeller speed. However,
none of these measures were successful in eliminating the tonal components completely. They seem to be
attributed to a large-scale stationary circumferential inflow disturbance due to a slightly asymmetric inflow
from the laboratory environment.
Fig. 14 shows the measured and the SEM-predicted overall sound power levels of the fan for a number of
operating points j (see Eq. (1)). Peaks at multiples of blade passing and impeller shaft frequency had been
removed in the underlying measured frequency spectra. One should not focus on operating points smaller than
j ¼ 0.17, because stall—which is not modeled here—more and more dominates the noise. Again, it is
remarkable, how much the noise in the ‘‘healthy’’ region of fan operation is determined by the inflow
turbulence. The combination TCS+HC+BLR yields the lowest noise levels which is believed to be more or
less the pure self-noise of the fan. The SEM predicts the measured tendencies correctly. The higher the
turbulence level, the better is the quantitative agreement. Obviously—and as expected—the more accurate
prediction of the self-noise requires refined models.

70

60

50
PSDLW [dB]

40

30

20

10

0
100 500 1000 5000 10000
(a) f [Hz]

70

60

50
PSDLW [dB]

40

30

20

10

0
100 500 1000 5000 10000
(b) f [Hz]

Fig. 13. Measured (___________) and SEM (- - - - - - - - -) predicted sound power spectra; (a) TCS+BLR, (b) TCS; ref 1012 W/Hz; tones at
multiples of BPF in the measured data are dashed.
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 67

105

100

95

OAPWL [dB]
90

85

80

75

70
0.00 0.04 0.08 0.12 0.16 0.20 0.24 0.28 0.32
ϕ

Fig. 14. Measured (peaks at multiples of impeller shaft frequency removed ) and SEM-predicted overall sound power levels (ref 1012 W)
as a function of the fan’s flow rate coefficient j; &: RPG2, n: NI, J: TCS+HC+BLR, filled plot symbols indicate the corresponding
SEM-prediction.

100

10-2
PSDF [N2/Hz]

10-4

10-6

10-8

10-10
100 1000 10000
f [Hz]

Fig. 15. LES-predicted unsteady impeller force (RPG2).

5.3. LES-predicted fan noise spectrum

As already mentioned the LES is exclusively applied to the inflow configuration RPG2. The discussion and
post-processing of the unsteady flow field data are based on a time interval of LES-predicted data which does
not include the physically non-relevant period during start up from the initial condition. In general, the time
interval considered corresponds to 10 revolutions of the impeller. The power spectral density is estimated by
dividing the discrete-time signal into seven 50%-overlapping sections, then applying a Hanning window and
averaging the spectra from all sections.
The unsteady overall impeller force is evaluated by a numerical integration of the unsteady pressures over
each blade surface element of the impeller. Fig. 15 depicts the power spectral density of the thus obtained force
PSDF.
Using Eq. (7) and substituting z PSDF by the overall impeller force gives the sound power spectrum as
shown in Fig. 16. The ‘‘erroneous’’ tonal components in the experimental data are dashed. In general, the
agreement between the LES predicted and the measured sound spectrum is reasonably good. Reasons for the
deviation at some frequencies are the inaccurate resolution of the blade boundary layer, the neglected tip
clearance, and, after all, the limitations of the acoustical model with respect to acoustic compactness.
 ARTICLE IN PRESS
68 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

80

70

60

PSDLW [dB]
50

40

30

20

10

0
100 500 1000 5000 10000
f [Hz]

Fig. 16. Fan sound power spectra (RPG2); measured (___________), SEM (- - - - - - - - - -) and LES ( ) predicted; tones at multiples of
BPF in the measured data are dashed; ref 1012 W/Hz.

With regard to the limits of the frequency resolution, the LES is able to resolve velocity fluctuations
downstream of the turbulence generator up to approx. 300 Hz with the numerical grid as described. However,
the radiated sound is predicted reasonably correctly far above 1000 Hz. This is due to the fact that the
convection velocity of the turbulent eddies over the involved solid surfaces varies roughly by a factor of 5. In
the turbulence generating mesh array, the convection velocity is in the order of the axial velocity in the
stationary frame of reference, whereas over the blade surfaces the eddies are convected with approximately the
relative velocity in the rotating frame of reference. The kinematics in low-pressure fans is typically such that
the axial velocity is much smaller than the circumferential and thus the relative velocity. Hence, assuming a
size of turbulent eddies which does not vary as the eddies are convected, the increase of convection velocity by
a factor 5 leads to a captured frequency range up to 1500 Hz in the force fluctuations on the blades and
subsequently in the acoustic spectra.

6. Conclusions

A special merit of the SEM is the fact that the spectral contribution to the overall spectrum from each
source is quantified separately. In order to predict the—in many industrial fan applications dominant—
‘‘turbulent ingestion noise’’, the intensity and the integral length scale of the ingested turbulence must be
given. This requires a ‘‘catalog’’ from which one can take these data for realistic inflow configurations. In this
study the turbulent inflow parameters have been determined experimentally for a variety of inflow
configurations. A hemispherical turbulent control screen surrounding the bellmouth-type inlet and other
measures have reduced the turbulence intensity of the inflow to the impeller to less than 1%. On the other
hand a grid-type ‘‘turbulence generator’’ provided more than 20% turbulence intensity. As already
demonstrated in previous studies from other authors the SEM is very easy to apply and requires very little
computational effort. Here, it predicts reasonably well the influence of the fan operating point and the nature
of the inflow on the noise spectra—without any further tuning of parameters. However, the predicted noise
spectra appear unrealistically ‘‘smooth’’, since the underlying empirical data is averaged and modeled in the
frequency domain. The SEM is not able to capture details of the fan geometry such as the form of the blade
profile or shape (e.g. skewed) or flow phenomena such as flow separation. A systematic drawback of the here
employed SEM is that it does not reproduce the cut-off of the broad-band noise spectrum; this inability to
predict properly sound from non-compact sources would have been palliated by finer analytical non-compact
models. Nevertheless, the SEM-predicted overall sound power levels are in good agreement with
measurements—thanks to averaging due to integration.
The LES allows to compute the acoustic sources, i.e. the fluctuating force on the blades, in the time domain
from first principles. Since the source prediction is based on the flow through the true geometry of the impeller
 ARTICLE IN PRESS
T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70 69

and the machine, the method yields inherently the influence of geometric details. So far the effect of the
ingested turbulence on the sound sources and the fan noise has been predicted surprisingly well. Details of the
source characteristics and their origin, averaged out in SEM, are obtained. Numerical experiments to predict
the pure self-noise, however, failed and are not reported here; in order to capture these effects a finer
resolution of the wall boundary layer is essential. This would require an increase of the size of the numerical
grid which easily can be beyond the scope of available computer capacity. Even for the reported case of highly
turbulent inflow, the computational cost are immense. Further improvements are expected by taking into
account additional geometric features such as the tip clearance which was omitted because of computational
resources. The acoustic results presented are based on a rather restricted acoustic model. Strictly speaking the
model is valid only for incoherent blade forces, i.e. without interaction or phase shift between the single
blades. Therefore, refined modeling of the sound field should also improve the accuracy of the predicted noise.

Acknowledgements

Parts of the work were supported by the ‘‘Bundesministerium für Wirtschaft über die Arbeitsgemeinschaft
industrieller Forschungsvereinigungen’’ Otto von Guericke ‘‘e.V. (AiF)’’ of the Federal Republic of Germany
and the German Research Foundation (DFG). The authors express their appreciation to C. Kato of the
University of Tokyo. Supported by the German Academic Exchange Service (DAAD) Prof. Kato hosted the
third author for more than a year in his research group. We gratefully acknowledge all this support.

References

[1] I.J. Sharland, Sources of noise in axial flow fans, Journal of Sound and Vibration 1 (3) (1964) 302–322.
[2] C.L. Morfey, The acoustic of axial flow machines, Journal of Sound and Vibration 22 (4) (1972) 445–466.
[3] L. Bommes, J. Fricke, R. Grundmann (Eds.), Ventilatoren. 2. Auflage, Vulkan Verlag, Essen, 2003 pp. 352–377.
[4] F.T. Brooks, D.S. Pope, M.A. Marcolini, Airfoil self-noise and prediction, NASA RP-1218, 1989.
[5] M.V. Lowson, Assessment and prediction of wind turbine noise, Flow Solutions Report 92/19, ETSU W/13/00284/REP, December
1992, pp. 1–59
[6] M.V. Lowson, S.P. Fiddes, Design prediction model for wind turbine noise: 1. Basic aerodynamic and acoustic models, Flow Solution
Report 93/06, W/13/00317/00/00, November 1993, pp. 1–46.
[7] M.V. Lowson, Theory and experiment for wind turbine noise, AIAA Paper 94-0119, 1994.
[8] P. Moriarty, P. Migliore, Semi-empirical aeroacoustic noise prediction code for wind turbines, Technical Report National Renewable
Energy Laboratory/TP-500-34478, 2003.
[9] R.K. Amiet, Acoustic radiation from an airfoil in a turbulent stream, Journal of Sound and Vibration 41 (4) (1975) 407–420.
[10] R.W. Paterson, R.K. Amiet, Noise and surface pressure response of an airfoil to incident turbulence, Journal of Aircraft 14 (8) (1977).
[11] J.E. Ffowks-Williams, L.H. Hall, Aerodynamic sound generation by turbulent flow in the vicinity of a scattered half plane, Journal of
Fluid Mechanics 40 (4) (1970) 657–670.
[12] R.K. Amiet, Noise due to turbulent flow past a trailing edge, Journal of Sound and Vibration 47 (3) (1976) 387–393.
[13] S.A.L. Glegg, Airfoil self noise generated in a cascade, AIAA Paper No. 96-1739, 1996.
[14] S.A.L. Glegg, C. Jochault, Broadband self noise from a ducted fan, AIAA Paper 97-1612, 1997.
[15] M. Roger, S. Moreau, M. Wang, An analytical model for predicting airfoil self-noise using wall-pressure statistics, Annual Research
Briefs, Center for Turbulence Research, Ecole Centrale de Lyon, 2002, pp. 405–414.
[16] G. Algermissen, R. Siegert, Th. Spindler, Numerical simulation of aeroacoustic sound generated by fans under installation
conditions, AIAA Paper No. 2001-2174, 2001.
[17] S. Kouidri, D. Fedala, T. Belamri, R. Rey, Comparative study of the aeroacoustic behaviour of two axial flow fans with different
sweep angles, ASME Fluid Engineering Summer Conference, Paper No. FEDSM2005-77424, Houston, TX, 2005.
[18] D. You, M. Wang, R. Mittal, P. Moin, Study of rotor tip-clearance flow using large-eddy simulation, AIAA Journal No. 2003-0838,
(2003).
[19] C. Kato, H. Mukai, A. Manabe, Large eddy simulation of unsteady flow in a mixed-flow pump, The Ninth International Symposium
on Transport Phenomena and Dynamics of Rotating Machinery, Honolulu, Hawaii, 2002.
[20] ISO 5136, Determination of sound power radiated into a duct by fans, 1990.
[21] J. Scoles, J.B. Ollerhead, An experimental study of the effects of an inlet flow conditioner on the noise of a low speed axial flow fan,
National Gas Turbine Establishment, Report No. AT/2170/049/XR, University of Technology, Loughborough, 1981.
[22] J. Scheimann, J.D. Brooks, Comparison of experimental and theoretical turbulence reduction from screens, honeycomb and
honeycomb–screen-combinations, Journal of Aircraft 18 (8) (1981).
[23] ISO 3745, Determination of sound power levels of noise sources using sound pressure—precision methods for anechoic and hemi-
anechoic rooms, 2000.
 ARTICLE IN PRESS
70 T. Carolus et al. / Journal of Sound and Vibration 300 (2007) 50–70

[24] J.O. Hinze, Turbulence, 2nd ed., McGraw-Hill, 1975.

[25] A.V. Smol’yakov, V.M. Tkachenko, The Measurement of Turbulent Fluctuations, Springer, Berlin, Heidelberg, New York, 1983.
[26] W.L. Keith, D.A. Hurdis, B.M. Abraham, A comparison of the turbulent wall-pressure spectra, Transactions of the ASME, Journal of
Fluids Engineering 114 (1992) 338–347.
[27] B.D. Mugridge, Broadband noise generation by aerofoils and axial flow fans, AIAA Paper 73-1018, 1973.
[28] C. Kato, M. Kaiho, A. Manabe, An overset finite element large eddy simulation method with application to turbomachinery and
aeroacoustics, Journal Applied Mechanics (70) (2003) 32–43.
[29] C. Kato, M. Ikegawa, Large eddy simulation of unsteady turbulent wake of a circular cylinder using the finite element method,
Advances in Numerical Simulation of Turbulent Flows, ASME FED 117 (1991) 49–56.
[30] H. Reese, C. Kato, Th., Carolus, Large eddy simulation of acoustical sources in a low pressure axial-flow fan due to highly turbulent
inflow conditions, Seisan Kenkyu, 57(1), January 2005, Tokyo, Japan.
[31] M. Schneider, Der Einfluss der Zuströmbedingungen auf das breitbandige Geräusch eines Axialventilators (Dr.-Ing. Dissertation
Universität Siegen, 2005), Fortschritt-Berichte VDI Reihe 7 Nr. 478. Düsseldorf: VDI Verlag 2006.
[32] P.E. Roach, The generation of nearly isotropic turbulence by means of grids, International Journal Heat and Fluid Flow 8 (2) (1986)
82–92.