Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

15th AIAA/CEAS Aeroacoustics Conference (30th AIAA Aeroacoustics Conference)

11 - 13 May 2009, Miami, Florida

AIAA 2009-3360

Measurement and Prediction


of the Tonal Noise of a Centrifugal Compressor at Inlet
Johanna Ingenito
Liebherr Aerospace Toulouse SAS, Toulouse, 31016, France

and Michel Roger

Ecole
Centrale de Lyon, Ecully,
69134, France.
The paper addresses two aspects of the tonal noise generated by a subsonic, centrifugal
compressor. The first aspect is an experimental investigation, based on a modal detection in the inlet duct, proving that the wake-interaction noise from the impingement of
the impeller blade wakes on the diffuser vanes is the dominant mechanism. The second
aspect is the analytical modelling of the sound transmission from the impeller outer radius
through the inter-blade channels and its recombination as helical modes in the inlet duct.
The propagation inside the inter-blade channels is described by a slowly-varying bent duct
approach, for which a simplified description of the channel geometry is considered. The
modal pressure patterns predicted analytically are compared with a numerical simulation
using a commercial software, and the discrepancies suggest that significant modal coupling
occurs, ignored in the analytical solution. The presented work is part of a global predicting
methodology in which different analytical models are chained from the sources to an external observer, successively dealing with the modal excitation of the impeller channels from
sources distributed on the diffuser, the sound transmission through the channels, the sound
recombination at inlet and the radiation by the inlet duct termination. More specifically
the coupling procedure between the inter-blade channel modes and the helical modes in
the air-inlet duct is discussed, as only an element of the analytical tool-box currently under
development.

Nomenclature
A0 (S)
B
c
D,D0
E
hs
k,K
M
N (S)
p

R
r(S)
s
Tj
U ,U0
V
(x, y)
PhD

modeshape function
blade number
speed of sound
density (non-dimensional)
Bernoulli constant
curvature parameter
wave number
axial Mach number
slowly varying normalization function
acoustic pressure
local radius (non-dimensional)
local curvature radius
curvilinear abscissa
Chebyshev polynomials
slowly variable mean-flow axial velocity
vane number
Cartesian coordinates
Student, Laboratoire de M
ecanique des Fluides et dAcoustique
Laboratoire de M
ecanique des Fluides et dAcoustique

Professor,

1 of 18
Institute of
Aeronautics
Copyright 2009 by the authors. Published by the AmericanAmerican
Institute of Aeronautics
and
Astronautics, and
Inc., Astronautics
with permission.

,
k

,
Subscript
m, n

small parameter characteristic of the multiple-scale solution


ratio of specific heats
fluid density
acoustic potential
azimutal location of the k th microphone
angular frequency
impeller rotational speed
modal orders

I.

Introduction

oise still remains a critical issue in aeronautics, not only for the environment of airports but also for the
N
working staff on the ground during the maintenance phase of an aircraft. In that case the main sources
of noise are the auxiliary power unit and the air conditioning units. For air conditioning systems, suppliers
like Liebherr Aerospace Toulouse (LTS) receive drastic recommendations from aircraft manufacturers and
are facing more and more stringent requirements. Most typical air system architectures designed until today
involved a transonic cooling fan taking air directly from outside through an inlet duct, whereas the other
elements of the system all mounted on the same shaft, namely the radial compressor and the turbine, had
no direct air path to the operating area of the staff. Many efforts towards the noise reduction of the fan have
been performed by LTS in the past. With the emergence of electrical aircraft, new air system architectures
are foreseen, for which the fan is suppressed and the centrifugal compressor inlet is directly connected to the
exterior. The compressor now most likely becomes the main source of noise during the maintenance phase.
As a result, additional efforts are required to understand the associated aeroacoustic phenomena.
The final objective of the supplier is to be able to pre-design a centrifugal compressor as quiet as possible at
a reasonable cost. A typical compressor is made of a radial impeller associated with a vaned diffuser, as shown
in Fig. 1. Due to the geometrical complexity of the impeller-vane-casing assembly, numerical simulations
are not tractable and analytical methods may be preferred, when available. The general approach chosen
here for compressor noise modelling is to try and reproduce all generating and propagating mechanisms from
the sources to an outside observer by analytical techniques. This choice implies simplifications in terms of
geometry and mean flow features, as well as an already identified source mechanism, assumed dominant. The
periodic impingement of the impeller wakes onto the vanes, called rotor-stator interaction, is assumed the
mechanism of interest here. This kind of source excites spiral waves of propagation at the impeller periphery,
as stated by a model addressed in a previous paper.1 When transmitted into the inter-blade channels of the
impeller these waves force slowly-varying modes inside the channels. Next the channel modes recombine at
the impeller inlet and give rise to helical modes that are transmitted along the inlet duct. Finally the helical
modes radiate outside by the duct termination. Some preliminary aspects of these chained mechanisms have
been discussed by Ingenito & Roger.2 Essentially a modal splitting of the waves in the different sections of
the sound path is justified by the relatively high frequencies of the sound: at the blade passing frequency
and its harmonics, the cross-sections of the inter-blade channels and of the inlet duct are non-compact. In
each section a specific model is derived, based on the orthogonal modes of propagation of the local geometry.
The adjacent sections are plugged in with each other by means of mode-matching equations. According to
this view each sub-model can be considered an element of a tool-box which is currently under development.
Within that scope, the present paper is mainly dealing with the tool-box elements related with the sound
transmission from the impeller interblade channels to the inlet duct. It is also aimed at confirming that
the wake-interaction noise generated in the outer part of the impeller and around the vanes of the diffuser
is the dominant contribution, as assumed in the modelling approach. This part of the work is achieved by
a dedicated experiment. The initial configuration of the compressor shown in Fig. 1 has been simplified
for a better assessment of intrinsic aeroacoustic sources, independently of installation effects. The inlet
plenum currently installed on a real aircraft is replaced by a straight duct for which the sound transmission
can be interpreted easily in terms of duct modes. A schematic view featuring the spinner and the inner
duct wall is given in Fig. 2. The experiment is first described and discussed in section II. Essentially a
modal detection is performed in the inlet duct, showing that wake-interaction noise indeed dominates the
tonal noise of the compressor. Section III is dealing with the slowly-varying duct approximation used for

2 of 18
American Institute of Aeronautics and Astronautics

the sound propagation inside the inter-blade channels. The theoretical background is Brambley & Peakes3
extension of Rienstras analysis to bent ducts. The multiple-scale approximation for bent slowly-varying
ducts proposed by Brambley & Peake3 for circular ducts has been adapted assuming that the channel has
locally a nearly rectangular cross-section.
Volute

1
Inlet duct
3

2
Diffuser

Impeller
Figure 1. Centrifugal compressor (courtesy of Liebherr-Aerospace Toulouse),
showing the main elements: 1, connection between the diffuser and the impeller;1
2, propagation inside the passage; 3, connection between the impeller and the
inlet.

1
2
3
M

Figure 2. Schematic view of the compressor blading, featuring the straight inlet
duct considered in the present modelling approach.

II.
A.

Experimental Modal Detection at Inlet

Setup and Processing Technique

The detection is not aimed at discriminating the upstream and downstream propagating waves but only
at extracting the azimuthal patterns from which the signature of the dominant generating mechanism can
be identified. Therefore a minimum instrumentation and the simplest available data-analysis technique are

3 of 18
American Institute of Aeronautics and Astronautics

implemented.

Figure 3. Picture of the inlet duct and of the rotating detection ring, featuring
the microphones and connections. The compressor and its surroundings are
hidden behind the insulating vertical plate seen in the background.

The modal detection is performed at isolated frequencies. The principle has been described by Sijtsma4
and is shortly outlined here. The acoustic pressure inside the duct can be expanded on the set of propagation
modes and its trace on the wall p() is expressed as a sum over the azimuthal modes by means of a Fourier
series
Z 2

X
1
i m
pm e
pm =
p() =
p() ei m d
2

0
m=

Therefore the modal amplitudes pm are obtained by performing a time FFT and next evaluating the integral
over a cross-section perimeter, for signals measured simultaneously. Experimentally they are estimated by
means of a discretized form of the integral, the instantaneous pressure being measured on a finite number
of microphones arranged as a circular array. Let k be the azimuthal location of the k th microphone and K
the total number of microphones. The estimation am of the complex modal amplitude pm at frequency f
reads
K
1 X
pk (f )eimk
am =
K
k=1

where pk is the fast-Fourier transform of the pressure signal. The mode-splitting technique would require
simultaneous measurements made on a large number of circumferentially distributed microphones, especially
if higher-order modes are needed. The alternative technique applied in this paper makes use of a small
number of microphones mounted on a rotating ring. Complementary measurements are made at successive
time steps with different azimuthal positions of the ring. The advantages are a lower cost related to the
number of microphones and a better trade-off between azimuthal resolution and implementation issues
due to microphone size, especially in a small-diameter duct. The drawback is that the time reference is
lost. Therefore a reference signal is needed, provided by a fixed microphone independent of the array of
moving microphones, and the transfer function with the reference signal must be considered. Furthermore
the rotational speed of the impeller cannot be maintained rigorously constant along the full acquisition

4 of 18
American Institute of Aeronautics and Astronautics

procedure. To solve this issue, the blade passing frequency clearly identified on the pressure signal FFT for
each position of the ring can be used afterwards to rescale the time axis.
If the index 0 stands for the signal measured by the fixed reference microphone at any acquisition, the
modal amplitude is expressed as
am =

K
1 X
1/2
pk p0 eimk / (p0 p0 ) ,
K
k=1

where the star denotes the complex conjugate.


The present experimental setup is made of one stationary, reference microphone and three moving microphones in a rotating ring, driven by a remote-control electric motor. The inner diameter of the duct is
10.6 cm. All microphones are nearly flush-mounted, in a very small cavity behind a pin-hole orifice of the
inner wall. The rotating ring and the microphone connections are shown in Fig. 3. The inlet duct is inserted
through a large insulating plate to keep the inlet free of possible contamination by regenerated sound from
outside. Practically the inlet duct termination is bevelled as a sharp edge. This makes the termination a
perfect scattering edge for which analytical models for the sound radiation such as Lordi & Homiczs are relevant,5 provided that the flow can be assumed uniform. No strong inflow distortion in the azimuthal direction
is expected if the inlet is well above the ground, so that the intrinsic tonal-noise sources of the compressor
dominate. The experiment is aimed at confirming the interaction of the wakes from the impeller trailing
edges with the diffuser vanes as the dominant mechanism. Indeed, even though the radiating properties of
that noise depend on the effective transmission in the impeller and on the recombination of helical modes in
the inlet, its modal structure can be shown to be reproduced identical in the inlet duct.2 Therefore, since
the wake interaction has a specific modal content imposed by the relative numbers of blades and vanes in
the impeller and the diffuser, an azimuthal mode detection at inlet is enough to conclude.
100

160

17,000 rpm

()

30,000 rpm

100

80

longitudinal
resonances

40

selfsimilar
broadband
loading noise

20

(1,0)
cutoff
frequencies of
duct modes

60

40

60

10

120

PSD 50 log

SPL ( dB ref. 2 105 Pa )

80

140

(2,0)

(3,0)

20
2

10

10

10

10

10

10

f / BPF

frequency (Hz)

(a)

(b)

Figure 4. (a): typical sound spectra measured by the flush-mounted microphones. Rotation speed 17,000 rpm (blue) and 30,000 rpm (black). Resonant
frequencies due to longitudinal reflections and duct-mode cut-off frequencies are
pointed for completeness. (b): scaling attempt featuring the self-similar range
and the non-similar broadband, high-frequency hump.

B.

Results

The investigated compressor has B = 17 blades and V = 19 vanes. At any multiple m of the blade passing
frequency, the most efficient mode is given by the lowest-order value of n = mB sV , s being any relative
integer, and corresponds to the highest rotational phase speed n = mB /n. At the blade passing frequency
(m = 1) which is of major interest for acoustic issues in view of the high rotational speed of the impeller,

5 of 18
American Institute of Aeronautics and Astronautics

the expected dominant modal order is n = 2. The presence of this mode in the modal detection will
therefore validate the relevance of the present model. Typical pressure spectra measured by a flush-mounted
microphone are plotted in Fig. 4-a for two different rotating speeds of the impeller. The higher speed of
30,000 rpm is close to the optimal design whereas the lower speed of 17,000 rpm is clearly off-design. Different
features are recognized. The blade passing frequency (BPF) and its first harmonic (2BPF) clearly emerge
from the broadband noise. The second harmonic 3BPF is also visible at 17,000 rpm. Small peaks at very
low frequencies (around 200 and 300 Hz), not related to the rotation speed, are attributed to longitudinal
resonances of the inlet duct. The broadband noise also exhibits sudden jumps followed by areas of monotonic
decrease at well defined frequencies, typically here 1880 Hz, 3114 Hz and 4282 Hz. These values are pointed
on the plots and correspond to the cut-off frequencies of the duct modes (1,0), (2,0) and (3,0). Since the
broadband noise sources force all possible modes and since a cut-off frequency is a resonant condition, such
jumps are expected. All aforementioned features suggest that the measurements are of acoustic nature, also
for what is dealing with the broadband noise, not investigated here. The BPF tone at 17,000 rpm is at a
very high level when compared to its level at the higher speed of 30,000 rpm. Normalized spectra are plotted
in Fig. 4-b based on the frequency made non-dimensional by the BPF and assuming the typical dipole-like
behaviour of loading noise to scale the sound level. The low-frequency part of the broadband spectrum
produces a nice collapse, suggesting that the flow features responsible for this signature are self-similar. In
contrast the high-frequency broadband hump behaves differently. As better seen in Fig. 4-a, the frequency
range of the hump is nearly independent of the rotating speed of the impeller, even though the corresponding
level increases with that speed.
45
40

0.05
35

0.25

0.06
17,000 rpm

30,000 rpm

17,000 rpm 2
BPF

0.2

1 BPF

1 BPF

0.04

30

0.15
25
0.03
20

0.1
15

0.02

10

0.05

0.01
5
0
10

5
0
5
azimutal mode order

(a)

10

0
10

5
0
5
azimutal mode order

10

0
10

(b)

5
0
5
azimutal mode order

10

(c)

Figure 5. Azimuthal mode spectra as deduced from the post-processing of pressure signals. Rotation speed 17,000 rpm for the blade passing frequency (a) and
its first harmonic (b), and 30,000 rpm for the blade passing frequency (c).

The modal spectrum of the BPF tone at a rotational speed of 30,000 rpm is plotted in Fig. 5-c. The
expected (-2) mode is clearly seen, with only residual modal amplitudes elsewhere (note that the secondary
modes selected by the interaction equation n = mB sV do not enter the range of investigated modes). This
shows that the impeller-diffuser wake interaction is the only contributing mechanism for this running point
of the compressor. The results at 17,000 rpm, plotted in Fig. 5-a, exhibit unexpected modal contributions
at n = 1 and 2. This off-design configuration possibly triggers flow instabilities which may induce modal
artefacts, and can be related to the apparent splitting of the BPF tone in two noticed in Fig. 4-a. More
precisely an extra tone is seen, slightly higher than the BPF tone for 17,000 rpm, the origin of which is
not known. However at twice the blade passing frequency the dominant mode is n = -4, again confirming
the wake interaction as the dominant mechanism. Finally the modal detection shows that the tonal noise
originates from the diffuser vanes and that predicting the sound transmission from the vanes towards the
inlet duct termination is a necessary task for the acoustic design.

6 of 18
American Institute of Aeronautics and Astronautics

III.

Propagation Inside an Inter-Blade Channel

Except for a phase-shift, the acoustic field is identical in each inter-blade channel of the impeller; therefore
the propagation only needs to be investigated in a single channel. The true geometry of a channel (see
Fig. 6(a)) involves a progressive curvature, torsion and cross-section variation. Furthermore the peripheral
end at outlet is an arc of a short cylinder and the inlet end is a sector of an annular disc. In the present study,
simplifications must be accepted for a more tractable modeling. First the torsion is neglected and the curved
mean line of the channel is assumed in a meridian plane, as shown in Fig. 6(b). Indeed the phase shifts
between adjacent channels, as well as the curvature leading from the radial orientation to the axial one, are
believed the most important features to include in the analysis. In addition, the cross section normal to the
curved channel axis is assimilated to a rectangle, and is supposed to vary from nearly a square at the inlet to
a narrow rectangle at the outlet. This is only valid in the limit of a large number of channels. The tangential
coordinate and the coordinate along the curvature radius can then be assimilated to Cartesian coordinates
to describe the cross section and cosines can be used as modal functions. Different approaches have been
then considered to propagate the sound inside the channel. The multi-modal approach proposed by Felix6
and dedicated to arbitrary varying ducts would be the most accurate. Its drawback is the need to calculate
the end impedance, which requires computing the acoustic pressure along the curvilinear abscissa by means
of a Runge-kutta algorithm with adaptive step. This method is quite CPU-time consuming for the present
needs and has not been retained for further integration into a global compressor model. Another approach
proposed by Brambley and Peake3 is based on an extension of the multiple-scale analysis first developed by
Rienstra7 and completed with a multiple-scale solution for cut-on cut-off transition by Ovenden8 to curved
ducts. This approach has been selected because it is very fast and fully compatible with the need to match
the solution with other propagation models inside the compressor. The methodology is presented below and
the analytical results will be compared in section 2 to numerical results obtained with ACTRAN/TM.9
Lx1
Ly1

Lx2

x
Ly2
(b) Simplified geometry

(a) Real geometry : with torsion

Figure 6. Inter-blade channel geometry : with (a) and without (b) torsion. The
section is close to a square at inlet and a narrow rectangle at outlet.

It is assumed that the rectangular cross-section varies slowly along the centerline of the channel. A crosssection is described in duct-centered Cartesian coordinates (x, y) in planes normal to the duct centerline
and with the curvilinear abscissa s along the centerline as shown in Fig. 6(b), s = corresponds to far
downstream towards the impeller outlet and the diffuser. The centerline curvature is 1/
r(S), r(S) being the
local curvature radius and S = s is the scaled coordinate along which the duct is slowly varying. The duct
inner and outer coordinates are xi,e (S) and the lateral boundaries are yi,e (S).
A.

Analytical Approach

The undisturbed velocity field inside the duct is defined as U = U es + V ex + W ey , the density is D, and
it is assumed that all mean-flow variables vary slowly along the duct: U = U0 + O(2 ), V = V1 + O(3 ),
7 of 18
American Institute of Aeronautics and Astronautics

W = W1 + O(3 ), where U0 (S, x, y) = U+ (S)/hs , with hs = 1 (S)x = 1 x/


r(S).
The quantity U+ (S) is found by equating the local mass flow rate at abscissa S to that far downstream,
to yield the implicit equation :
Z xe Z ye
U0 (S, x, y) D0 dx dy = U D (xe () xi ()) (ye () yi ()).
(1)
xi

yi

In reference papers on the multiple-scale analysis,7, 3 the duct is not moving and the flow equations are
written in a stationary reference frame. In the case of the rotating inter-blade channel the equations must
be written in a reference frame attached to the channel and describe the relative fluid motion. As such they
have to include inertial effects associated with centrifugal and Coriolis forces. This is not strictly compatible
with the necessary assumption of a potential motion. More precisely the Coriolis acceleration points in the
tangential direction, normal to the simplified mean-flow gradients, which leads to a contradiction in the
equations. In practice, the complementary acceleration is responsible for the structuration of secondary
flows in the inter-blade channels. For simplicity, the secondary flows are ignored here and the Coriolis term
is discarded from the analysis. It can be shown that the net effect is just to introduce an additional term in
the Bernoulli equation for the leading-order terms, which now reads


2R
2 1/(1)
U02
D0 = ( 1) E
2
is the non-dimensional rotating speed of the impeller, R
the radial location in the impeller and E
where
7
the Bernoulli constant introduced by Rienstra.
Now superimpose a small disturbance of time dependence eit to the mean flow (U, D, P ), say (u, , p)
with u = . Again for the sake of making the problem tractable, the linearized equations for the disturbance
potential are written ignoring the complementary acceleration terms, both due to Coriolis and centrifugal
effects. The rotation is recognized to have an effect on the mean flow but it is assumed that it does not affect
significantly the sound propagation with respect to the mean flow. Further investigation could be needed to
justify such an approximation. The resulting equation is


D
1
1 D
. (D) = 0
(2)
2
Dt C Dt
D
where D/Dt = /t + U. is the derivative with respect to the mean flow. Assuming rigid walls, the
boundary conditions are : /x = 0 at x = xi,e and /y = 0 at y = yi,e .
The slowly-varying duct geometry is accounted for according to the multiple-scale approximation7 by
writing the potential as
(
)
Z


i S
2

= A0 (S, x, y) + A1 (S, x, y) + O( ) exp it


k(S )dS .
(3)

in which is the small parameter characteristic of the multiple-scale solution. Keeping just the O(1) terms
in equation 2 leads to



  2


A0
A0

k2

1
hs D 0
+
hs D 0
+
(4)
2 A0 = 0
hs D0 x
x
y
y
C02
hs
with = kU0 /hs . The O(1) boundary conditions are
A0
= 0 on x = xi,e (S)
x

and

A0
=0
y

on y = yi,e

(5)

The equation is separable for the variable y, which has the great interest of reducing the subsequent
numerical effort. Therefore the solution can be found as


m
A0 (S, x, y) = cos
(y yi ) A0 (S, x)
ye yi
8 of 18
American Institute of Aeronautics and Astronautics

where the function A0 (S, x) must satisfy the equation


"
2 #

2 A0 (S, x) (S) A0 (S, x)
k2
2
m

=0

2
+ A0 (S, x)
x2
hs
x
C02
hs
ye yi
The boundary condition is now :

A0
= 0 on x = xi,e (S)
x

(6)

(7)

The function A0 is normalised as


Z

A0 (S, x, y) = N (S)A0 (S, x, y)

xe
xi

ye

yi

U0 D0 2
A0 dx dy = 1.
C02

(8)

where N (S) is a slowly varying function.


The unknown amplitude N (S) is determined from the solvability condition obtained using the O() terms
from (2). After a great deal of algebra it is shown that the quantity F (S)N 2 (S) must be constant along the
duct, which for cut-on modes can be interpreted as conservation of energy.
The F (S) function is :



Z xe Z ye
U02
k
2 U0

1 2
+
dx dy
(9)
D0 A0
F (S) =
C02
hs
C0
xi
yi
With this numerical approach, we now have the leading-order solution (8) for the unsteady flow, in which
the local axial wavenumber and modeshape are determined by the equations (4) & (5) and the slowly varying
amplitude is then given by F (S)N 2 (S).
B.

Numerical Solution

Equations (6) & (7) can be solved numerically to determine the axial wavenumber k(S) and the corresponding
wave function A0 (S, x). The leading-order equation for A0 and k is written as the generalized eigenvalue
problem



!
!
!
U02
2U0
1

A
L 0
A0
2
2
2
0
C0
,
(10)
= k hs C0 hs
B0
0 1
B0
1
0
where

LA0

"
2 #

2
2 A0
(S) A0
m

=

+ A0
x2
hs x
C02
ye yi

(11)

and
B0 = kA0

(12)
10, 11

This problem can be solved numerically using a pseudo-spectral method with Chebyshev polynomials.
Let nx be the required number of collocation points in the x direction. The interval x [xi , xe ] is
mapped linearly onto the interval (x) [1, 1]. The Chebyshev polynomials are defined on [1, 1] by
Tj (cos ) = cos(j), with collocation points j = cos1 [j/(nx 1)]. Performing a change of variable, the
numerical domain is made of points xj such that


xi + xe
j
xi xe
1
xj =
(13)
+
cos
2
2
nx 1
The values found numerically will be the values of the function under consideration at these points,
bj = f (xj ), and will represent the unique interpolating function of the form
f (x) =

nX
x 1

aj Tj (x)

j=0

9 of 18
American Institute of Aeronautics and Astronautics

(14)

Furthermore, the multiplication of two functions is simply made by multiplying their stored values pointwise, while differentiation by x is represented by the matrices Dx , where
Dx(l,j) =

2
Alj
xe xi

(15)

with
Alj
Ajj
A00

cl (1)l+j
cj l j
j
=
2(1 j2 )

if
if

l =j
1 l = j nx 2

2(nx 1)2 + 1
= A(nx 1)(nx 1)
6

where cn = 2 if n is 0 or nx 1, and cn = 1 otherwise.


The system is now discretized for a rectangular duct by requiring the boundary conditions /x = 0 to
be satisfied at collocation points 0 and nx 1, and equation (10) to be satisfied at collocation points j for
j = 1, . . . , nx 2.
Some simplifications are possible. Let L be the matrix obtained by discretizing L at the interior collocation points, and let G1 and G2 be the matrices representing the multiplication by (2U0 )/(hs C02 ) and
(1 U02 /C02 )/h2s at the interior collocation points. Let Dbound be the matrix evaluating the derivatives on
the boundaries, derived from Dx . Then the discretized problem may be written

!
!
0
0
Dbound 0

bj
bj

(16)
L
0 + = k G1 G2 + ,
bj
bj
I
0
0
I
|
{z
}
{z
}
|
F

+
where bj are the values of A0 at the j collocation points, and b+
j = kbj . The values of bj are only needed at
the interior points, and not at points where the boundary condition is being enforced. Then, the first column
is nx wide, the second column is nx 2 wide, the first row is 2 deep, the second row is nx 2 deep, and the
third and fourth rows are nx 2 deep. bj contains one value for every collocation point, and b+
j contains one
value for every interior collocation point. The problem is solved thanks to eig Matlab functions.

C.
1.

Results and Comparisons


Mean Flow Effect

Typical results for the modal pressure patterns inside an inter-blade channel are shown in Fig. 7 for modes
propagating downstream from the inlet towards the diffuser and in Fig. 8 for modes propagating upstream
from the diffuser towards the inlet. The plotted quantity is the instantaneous pressure without flow and with
a mean flow of Mach number at the inlet M = 0.7. More precisely the Mach number is defined far away
upstream by a uniform value of 0.7 in the axial continuation of the inlet. At the fixed frequency of interest,
the convection effect by the mean flow is clearly identified. A contraction of the wavefronts is observed
when going from the Fig. 8(a) to the Fig. 8(b), as expected for upstream propagating waves. In contrast for
wave propagation in the direction of the mean flow, the wavefronts are elongated (Fig. 7(a) vs. Fig. 7(b)).
Furthermore, the effect of a non-zero Mach number is to shift the cut-off frequencies to lower values, so
that cut-off modes with no flow might now be allowed to propagate. An interesting feature is observed at
the high Mach number of M = 0.7. The pressure fluctuations tend to concentrate near the outer-casing
wall when approaching the inlet (see in Fig. 8(b) and 8(d)). This might be a refraction effect due to the
mean-velocity gradient in the transverse direction. Indeed, as shown in Fig. 9(a), the velocity is higher at
the casing wall and lower at the spinner according to the present model. Sound refraction across such a flow
deviates upstream propagating waves towards the outer wall and downstream propagating waves towards
the inner wall.

10 of 18
American Institute of Aeronautics and Astronautics

(a) Mode (0,0) - M0 = 0

(b) Mode (0,0) - M0 = 0.7

(c) Mode (0,1) - M0 = 0

(d) Mode (0,1) - M0 = 0.7

Figure 7. Downstream propagating modes from the inlet without (M0 = 0)/with
mean flow (M0 = 0.7) - f = 24000 Hz - m = 0 - = 33000 tr/min

11 of 18
American Institute of Aeronautics and Astronautics

(a) Mode (0,0) - M0 = 0

(b) Mode (0,0) - M0 = 0.7

(c) Mode (0,1) - M0 = 0

(d) Mode (0,1) - M0 = 0.7

Figure 8. Upstream-propagating modes from the diffuser without (M0 = 0) or


with mean flow (M0 = 0.7) - f = 24000 Hz - m = 0 - = 33000 tr/min

(a) Analytical mean flow

(b) Actran mean flow

Figure 9. Typical distributions of the mean flow velocity according to the analytical solution (a) and the ACTRAN software. Inlet Mach number M = 0.2 at
infinity.

12 of 18
American Institute of Aeronautics and Astronautics

2.

Comparison with a Commercial Software

Accompanying computations have been performed using the ACTRAN/TM software.9 Note that the selected
boundary conditions may not be physical because they have been chosen compatible with the analytical
assumptions previously stated. A single mode with unit amplitude is imposed at the downstream end of
the channel (outer radius of the impeller) and the upstream-travelling waves are totally absorbed at inlet.
Note also that the mean velocity field of the numerical simulation is different from the analytical model as
shown in Fig. 9. This artefact is attributed to different assumptions made in the ACTRAN computations.
It is expected to induce strong effects on the propagation, especially at high mean-flow Mach numbers. No
correction has been attempted yet, though needed. A further step will be to impose the analytical mean
flow in the simulations.
The comparison with the analytical model is discussed here for the incident plane-wave mode (0,0) at
outlet at Mach number 0.2. The slowly varying-duct mode solution is plotted in Fig. 10-a and the higherorder mode (0,1) is also plotted for the sake of the discussion. The numerical solution is shown in Fig. 11-a.

(a) Analytical mode (0,0)

(b) Analytical mode (0,1)

Figure 10. Instantaneous acoustic pressure patterns for upstream incident modes
injected at outlet. Inlet Mach number M0 = 0.2. (a): mode (0,0); (b): mode (0,1).
Amplitudes relative to an arbitrary incident pressure at outlet

The numerical results suggest that modal conversions occur along the channel because of the changing
geometry. Indeed the transverse variations cannot be explained by a single mode, and distortions are seen
in the streamwise direction with respect to what would be expected from the wavelength of a single mode.
In view of the pressure pattern of Fig. 10-b, the mode (0,1) is cut-off in the oulet rectangular cross-section
where it is forced but goes cut-on farther upstream due to the enlargement of the channel in the plane of
the figure, in such a way that the initial exponential damping cannot be effective. Therefore it must be
considered as a possibly contributing mode away from the oulet cross-section. The transition frequency
of the mode (0,2) takes place much closer to the inlet and this mode (not shown in the figure) cannot be
sustained by an excitation at outlet. But again it is an acceptable local solution at inlet. The weakness of
the analytical model is that at the present stage it cannot account for modal conversions, because the solving
procedure focusses on one mode and follows its modifications along the curved axis of the channel. Such
conversions are expected at the locations of cut-on to cut-off transitions along a bent duct. As a preliminary
and indicative attempt the analytical modal patterns of modes (0,0), (0,1) and (0,2) have been superimposed
with arbitrary relative amplitudes and phases, leading to the pressure map of Fig. 11-b. Even though the
result does not coincide with the numerical solution, some features such as the transverse skewness of the
wavefronts and different amplitude concentrations are pointed out. The mode (1,1) could have also been
considered since it transitions somewhere between the inlet and the outlet. But no better comparison has
been sought because of the lack of guideline for the determination of the relative amplitudes and phases of
the modes in the analytical solutions. Furthermore, part of the discrepancies are attributed to the differences
of mean-flow speed distribution, certainly not negligible even at a Mach number of 0.2.
Similar calculations at the higher Mach number of 0.5 are reported in Fig. 12. Though the concentration
13 of 18
American Institute of Aeronautics and Astronautics

(a) ACTRAN computation

(b) Analytical calculation

Figure 11. Colour maps of instantaneous acoustic pressure for the plane-wave
mode injected at outlet. Mach number M0 = 0.2. (a): Actran software ; (b):
present analytical model with an arbitrary superposition of the modes (0,0),
(0,1) and (0,2)

(a) ACTRAN computation

(b) Analytical computation

Figure 12. Colour map comparison of Acoustic pressure for plane wave injected
at mach flow M0 = 0.5 with Actran software (a) and with the present analytical
model (b) (single mode)

14 of 18
American Institute of Aeronautics and Astronautics

of the pressure fluctuations at the wall is observed in both predictions, the same overall discrepancies of the
pressure patterns are found, again probably due to the modal coupling, or to the difference of the mean flow.
Finally very different results are obtained, attributed to the different methodologies.
In the Brambley and Peakes approach,3 the analytical calculation consists in following the variation
of the wave number of a given mode all along the duct; no coupling between modes is considered.
In ACTRAN computations, due to the exact treatment of the cross-section variations and of the mean
curvature, modes different from the excitation are automatically generated inside the duct.
It is concluded that for strongly bent ducts the modal coupling is significant and cannot be ignored. The
question of whether or not the slowly-varying duct approximation can accomodate with a proper coupling
procedure still remains to be answered and requires further theoretical developments at the transitions.

IV.

Coupling Strategy

As indicated in the introduction, the coupling procedure is the last step to achieve before having an
integrated model. This model will allow getting a transfer function of the acoustic propagation inside the
centrifugal compressor and then to select the most relevant compressor design. It should be emphasized that
all the sub-models chosen in the frame of that study follow a modal approach which is an important criterion
to make the coupling easier.
The present study addresses the upstream acoustic propagation from the sources, up to the radiation by
the inlet duct termination (Fig. 13). This propagation inside the complete compressor is modelled separately
in each sub-part of the compressor using the simplified geometries described in Fig. 13 with:
Fig. 13(a): Connection between the diffuser and the impeller1
Fig. 13(b): Propagation inside the passage
Fig. 13(c): Connection between the impeller and the inlet2
Fig. 13(d): Propagation inside the straight inlet duct including the spinner2
Though the complete chaining will lead to a substantial complexity, resorting to a specific sub-model for
each part of the sound transmission path allows understanding clearly the underlying physics, including the
contribution of cut-on and cut-off modes, mode transitions and conversions, cascade effects on the reflection
and the transmission of waves, and so on. In its final form the prediction tool will help to define the most
suitable geometry, blade-and-vane counts and global parameters for a minimum noise, with the constraint
of imposed aerodynamic performances.
The coupling between the sub-models is a priori performed using a cascade process: each model solves
the continuity equations of an interface and the output of a first model is taken as the input of the next
one. But it must be emphasized that each connection generates additional reflected and transmitted waves.
As a result each interface includes four different sets of waves, propagating upstream and downstream
on both sides, whereas an isolated interface is generally investigated assuming only three set of waves,
namely a specified incident wave and the reflected and transmitted waves derived by solving the matching
equations. The chaining requires to re-address the equations by adding the fourth set of waves. For instance,
the recombination at the inlet of the impeller has been analyzed previously assuming a single upstream
propagating mode in each inter-blade channel as incident wave.2 The solving procedure provides the reflected
modes inside the channels and the helical modes transmitted in the inlet. The latter radiate outside by the
duct termination and produce back-reflections wich will be diffracted again at the impeller inlet. In the
same way, secondary upstream reflections are generated when some of the reflected channel modes at the
inlet reach the impeller periphery. Their complex amplitudes must be added in the definition of the incident
waves on the inlet interface, as additional unknowns which are in fact determined by the matching equations
of the complementary outlet interface. All these modifications do not alter the solvability of the matching
equations but only increase the size and/or the content of the matrices to be inverted. For a better tracking
of the waves, the equations can also be solved iteratively, as follows. A first prescribed wave on an interface
ignoring the second interface is used as an initial condition. First-approximation reflected and transmited
waves are determined, from which the secondary reflected waves are deduced (solving the equations at the
15 of 18
American Institute of Aeronautics and Astronautics

Lx1

M3

Ly1

RT

M2

Lx2

s
Ly2
(a)

(b)

R2

M1

R1

M1

M1

L
(d)

(c)

Figure 13. Modal propagation in the simplified geometries assumed in the study
(thin red arrows are taken into account in the present version of the analytical
model)

16 of 18
American Institute of Aeronautics and Astronautics

second interface). These waves are considered to re-define the incident conditions, and so on. Such an
iterative procedure is expected to converge since it reproduces the progressive onset of waves that would
physically follow a source switch-on. For a first insight, approximations can be accepted. Typically, higherorder modes of a cylindrical duct are known to be significantly reflected at the duct termination only just
above their cut-off frequency, whereas otherwise they are almost totally transmitted outside. Therefore the
reflected waves at the inlet duct end (indicated by thick arrows in Fig. 13) can be neglected as small when
compared to te transmitted ones (thin arrows in Fig. 13).
Another issue in the chaining is that the modal bases on both sides of a connection may not coincide,
essentially because of the simplifications introduced in the geometry. This occurs when connecting the
propagation inside the inter-blade channels (Fig. 13(b)) and the propagation in the inlet duct (Fig. 13(c)).
The channel modes are cosine-cosine functions chosen for convenience (see section III), whereas the true
modal shapes should be given by Bessel functions in the radial direction. The Bessel functions are more
suited for the matching with the inlet duct. A simple way to overcome this shortcoming is to expand the
radial cosine modes onto the Bessel-functions basis. Though not shown here the expansion has been proved
to be free of significant errors. Except for the first Bessel function, the projection between the two bases
needs around ten coefficients to be relevant.
The chaining between Fig. 13(a) and Fig. 13(b), as well as Fig. 13(c) and Fig. 13(d) remains to be
performed. When it is available, the resulting complete model will allow comparing the radiated acoustic
fields of different compressor designs and selecting the quieter one.

V.

Concluding Remarks

The present paper was focussed on some aspects of an analytical strategy developed by LTS to predict
the sound generation and transmission from the diffuser to the inlet duct of a vaned radial compressor. The
global model is formulated as different chained sub-models that can be solved analytically using a modal
approach. An experimental modal detection, based on an azimuthal investigation only, has been first carried
out at inlet to validate the main assumption made in the analytical approach, and has confirmed that the
wake-interaction noise due to the impingement of the impeller wakes on the diffuser vanes dominates the
tonal noise of the compressor. The emphasis has been on the propagation inside the inter-blade channels.
For this part of the work, the multiple-scale approximation for bent slowly-varying ducts initially proposed
by Brambley & Peake3 for circular ducts has been selected. The formalism has been adapted assuming
that the channel has locally a nearly rectangular cross-section. Some analytical results have been compared
with a commercial software providing an exact solution. The discrepancies are attributed to a significant
modal coupling inside the channels, ignored in the analytical solution. The complementary sub-models were
addressed in previously reported studies. All are considered as elements of an analytical tool-box currently
under development. The principle of the chaining has been described. The next step will be to implement
all the modifications resulting from the association of the different sub-models, typically the secondary
reflections. The global integrated tool is aimed at providing a transfer function between some excitation at
the impeller outer radius and the corresponding response in the inlet or outside. The latter will be used for
the preliminary design of low-noise compressors.

Acknowledgments
The study has been supported by Liebherr-Aerospace Toulouse.

References
1 Roger, M., Analytical modelling of wake-interaction noise in centrifugal compressors with vaned diffusers, AIAA Paper ,
, No. 2004-2994, May 10-12 2004.
2 Ingenito, J. and Roger, M., Analytical Modelling of Sound Transmission Through the Passage of Centrifugal Compressors, AIAA Paper , , No. 2007-3704, May 21-23 2007.
3 Brambley, E. and Peake, N., Sound Transmission in Strongly-Curved Slowly-Varying Cylindrical and Annular Lined
Ducts with Flow, AIAA Paper , Vol. 2006-2582, 2006.
4 Sijtsma, P. and Zillmann, J., In-Duct and Far-Field Mode Detection Techniques, AIAA Paper , , No. 2007-3439, May
21-23 2007.
5 Lordi, J., Homicz, G., and Rehm, R., Effects of finite duct lenght and blade chord on noise generation by a rotating

17 of 18
American Institute of Aeronautics and Astronautics

blade row, AIAA Paper , , No. 74-555, 1974.


6 Felix, S., Propagation acoustique dans les guides dondes courbes & probl`
eme avec source dans un
ecoulement cisaill
e,
Ph.D. thesis, Ecole Doctorale de lUniversit
e du Maine, France, 2002.
7 Rienstra, S., Sound transmission in slowly varying circular and annular lined ducts with flow, J. Fluid Mech, Vol. 380,
1999, pp. 279296.
8 Ovenden, N., A uniformly valid multiple scales solution for cut-on cut-off transition of sound in flow ducts, J. Sound
Vib., Vol. 286, 2005, pp. 403416.
9 Actran - Users Manual 2006 Aeroacoustics solutions : Actran/TM and Actran/LA. Version 1th of December 14, 2005 .
10 Canuto, C., Hussaini, M., Quarteroni, A., and Zang, T., Spectral Methods in Fluid Dynamics, Springer-Verlag, 1988.
11 Boyd, J. P., Chebyshev and Fourier Spectral Methods, Second Edition, DOVER Publications, Inc., 2000.

18 of 18
American Institute of Aeronautics and Astronautics

You might also like