Professional Documents
Culture Documents
A Control Scheme For Tuning
A Control Scheme For Tuning
Nomenclature
1 of 14
B. Control Objective
The aim of the control scheme to be presented in this work is to expose a system to acoustic boundary
conditions which are different than the actual ones. In the following, the set-up depicted in Fig. 1 is
considered. The system is connected to a duct, in which only plane waves propagate, terminated with an
acoustic boundary condition Z1 . Here, Z denotes the acoustic impedance, a complex valued function of
frequency. The acoustic impedance is defined as
Z = p̂/û, (1)
where p̂ and û denote Fourier transforms of acoustic pressure (which is considered to be scaled by the
characteristic impedance ρc, ρ being fluid density and c the speed of sound) and particle velocity normal to
Figure 1. System with different acoustic boundary conditions induced by a change in geometry or by imple-
mentation of a liner
2 of 14
p = f + g, (2a)
u = f − g. (2b)
Note here that the linear relations expressed in Eqs. (2) hold in frequency as well as in time domain.
In terms of the Riemann invariants, the acoustic boundary condition is expressed as the reflection coef-
ficient, which represents the ratio of the reflected and the incident wave. The reflection coefficient is related
to the impedance by
Z −1
R= . (3)
Z +1
Therefore, equivalently to imposing a certain desired impedance, the control scheme will be designed so as
to mimic a certain reflection coefficient.
Rcl f
SYSTEM
Rol g
Figure 2. System with controlled boundary conditions. The wave decomposer D extracts the downstream
propagating wave f from multiple pressure measurements p. Based on f , the controller K generates the
actuation command e. The uncontrolled reflection coefficient Rol is manipulated by the control scheme and
the actuator so that the system is exposed to a reflection coefficient given by Rcl
3 of 14
If R and G are assumed to be known from the frequency response measurements, the control law K can be
formally calculated from Eq. (6) if a desired closed-loop reflection coefficient Rcl is prescribed:
Rcl − Rol
K= . (7)
G
In order to implement the control scheme on a digital control board, a model for the control law K has to
be found. In the present work, this is done by applying a frequency domain system identification algorithm
(see Gustavsen and Semlyen11 ) to the discrete frequency response data, which is obtained from Eq. (7).
where τ ± = ∆x/(c ± u0 ). In these equations, g can be eliminated by evaluating Eq. (9) at t − τ − and using
the result in Eq. (8). This gives
from which it is concluded that the f wave can be identified in real time by using the scheme shown in Fig. 3
(right).
4 of 14
One drawback of this approach is that the feedback loop in Fig. 3 (right) is only marginally stable. To
see this, Eq. (10) is solved for fˆ in frequency domain:
−
p̂1 − p̂2 e−iωτ
fˆ = . (11)
1 − eiω(τ − +τ + )
The purely real eigenvalues are located at
2πn
ωn = , n = 0, ±1, ±2... (12)
τ+ + τ−
Note that this relation can be recast in the form
kn ∆x = πn(1 − M 2 ), (13)
which was given, e.g., by Åbom and Bodén12 in the analysis of the Two-Microphone-Method (TMM) in
frequency domain. It is important to note that, as the above authors pointed out, not only are the frequencies
given in Eq. (12) excluded from data processing, but the error in the TMM increases dramatically when
approaching theses frequencies. Also, decreasing the microphone spacing to move ω1 to higher frequencies
is not an option since, in this case, the accuracy of the TMM significantly deteriorates at low frequencies.12
In practice, this means that the identification algorithm cannot be used close to frequencies for which
Eq. (12) holds. Moreover, oscillations can be excited that would have to be suppressed with suitable notch
filters.
p1 p2 pn-1 pn
…
f
U0
g
x1 x2 … xn-1 xn
In the frequency domain, this issue is resolved by using the Multi-Microphone-Method.9, 10 Here, several
microphones at different axial locations, corresponding to unequal propagation delays, are used. This resolves
the issue of singularities at frequencies given by Eq. (12). The wave decomposition is then obtained from a
least squares fit to the measured complex pressure amplitudes. Another advantage of the MMM is that by
averaging over several microphones, flow noise is, at least partially, suppressed. This is even more important
in a real-time application where noise rejection techniques, such as cross-correlations with the excitation
signal, cannot be used. However, the MMM is applied in a post-processing step, where pressure records for
a certain time interval are available. For the impedance tuning control scheme, the decomposition has to be
accomplished in real-time. Therefore, application of the traditional MMM is not straightforward. However, a
time domain scheme, having similar advantages like the MMM in frequency domain (i.e., using more pressure
5 of 14
Equations (15) are now considered as an overdetermined linear system for f (t) and g(t), whose least squares
solution takes the form
p1 (t)
− + −
" # " #
f (t) 1 −1/(n − 1) . . . −1/(n − 1) p (t − τ1,2 ) − f (t − τ1,2 − τ1,2 )
= 2 .. . (16)
g(t) 0 −1/(n − 1) . . . −1/(n − 1)
.
− + −
pn (t − τ1,n ) − f (t − τ1,n − τ1,n )
A feedback scheme similar to the one shown in Fig. 3 can be constructed, mapping the n pressures to
the downstream traveling wave. It is, however, more instructive to consider the frequency domain solution
for fˆ, which is obtained from (16) as
− −
1
p̂1 − n−1 p̂2 e−iωτ1,2 + . . . + p̂n e−iωτ1,n
fˆ = + − + −
. (17)
1
1 − n−1 e−iω(τ1,2 +τ1,2 ) + . . . + e−iω(τ1,n +τ1,n )
+ −
The denominator in Eq. (17) has real zeros only where ω(τ1,k +τ1,k ) mod 2π = 0 for all k. If the propagation
delays associated with the microphone spacings are incommensurate, this will happen only at large frequen-
cies (except at ω = 0). Apart from this, all poles lie in the upper half of the ω-plane. A plot of the first few
poles of Eqs. (11) and (17) for the case n = 3 is shown in Fig. 5. Accordingly, a state space realization of
1.2
1
=(ω)(τ + + τ −)
0.8
0.6
0.4
0.2
0
−0.2
0 π 2π 3π 4π 5π 6π 7π 8π 9π
<(ω)(τ + + τ −)
Figure 5. Poles of the two-microphone identification scheme (Eq. (11), squares) and a three-microphone method
+ − + − √
(Eq. (17), circles) with (τ1,1 + τ1,1 )/(τ1,2 + τ1,2 ) = 2
6 of 14
A. Experimental Set-up
The experimental set-up considered is the acoustic test facility depicted schematically in Fig. 6. Seven 1/4”
microphones were installed at different axial positions. All microphones were used for frequency domain
model identification but only three of them for time domain wave decomposition in the control scheme.
The duct is made of aluminum with a wall thickness of 10 mm and has a diameter of 140 mm. Accordingly,
the cut-on frequency for the first non-planar mode is at f10 = 1435 Hz at ambient conditions.2 The upstream
end was equipped with an anechoic termination, the downstream end was open and unflanged. Speakers are
laterally mounted to the duct, close to the up- and downstream ends. For some experiments, an additional
low-reflecting end was mounted at the downstream end.
p e
DSP
anechoic end
open end
G
Rcl
flow Rol
reference plane
Figure 6. Experimental set-up used for the application of the control scheme
B. Model identification
As shown in Sec. II, the uncontrolled reflection coefficient Rol and the actuator transfer function G have to
be known to calculate the control law. Both can be determined from frequency response measurements of
the plane wave field. This procedure is explained in the following.
To identify the end element’s ĝ response according to Eq. (4), two independent acoustic states are
required. The first is generated by upstream acoustic forcing. In this case, ê = 0 and Rol can be obtained by
calculating the complex wave ratio ĝ/fˆ. A second state can now be realized by either only downstream or
simultaneous up- and downstream excitation. Since Rol is already known from the first state, the actuator
transfer function can be computed from Eq. (4) as
ĝ − Rol fˆ
G= . (18)
ê
The Riemann invariants fˆ and ĝ cannot be measured directly. However, as mentioned in Sec. III, they
can be estimated from frequency response measurements at multiple axial locations along a uniform duct.
For the test facility considered in this work, all 7 pressure measurements were used for frequency domain
wave decomposition.
7 of 14
1 1
0.8 0.8
|R|
0.6 0.6
|G|
1−M
1+M
0.4 M =0 0.4
M = 0.05 M =0
0.2 0.2 M = 0.05
Levine & Schwinger
0 0
-π -π
π π
-2 -2
arg G
arg R
0 0
π π
-2 -2
-π -π
200 300 400 500 600 Hz 800 200 300 400 500 600 Hz 800
frequency frequency
Figure 7. Reflection coefficient (left) and actuator transfer function (right) with and without mean flow
It may be desirable to model the open-loop reflection coefficient and/or the actuator transfer function
analytically (see, e.g., Venugopal & Bernstein14 or Pota and Kelkar15 ) and, in this way, be able to design the
controller completely a priori. However, to achieve a reasonable result, Rol and K have to be known quite
accurately, and it seems doubtful that a completely theoretical model will suffice, except for the simplest
configurations.
Model set-up
The system considered in this work (see Fig. 6) is modeled as a network of acoustic elements. The structure
of the network model is shown in Fig. 8. Individual elements are described by 2 × 2 linear time-invariant
input-output relations (so called acoustic two-ports). Plane wave modeling is sufficient since only frequencies
are considered for which all higher modes are cut-off.
An interior element is sufficiently described by a 2×2 mapping of acoustic variables. There are various
representations of this mapping but only the scattering form will be considered here. The scattering matrix
of an acoustic element relates the scattered and reflected waves to the incident waves in frequency domain
" # " #" #
fˆd S11 S12 fˆu
= , (19)
ĝu S21 S22 ĝd
where subscripts u and d denote up- and downstream positions, respectively. An equivalent time domain
8 of 14
duct
duct
duct
G duct
G
R R
representation has the form of a general state space system with two inputs and two outputs
" #
f
ẋ = Ax + B u ,
gd
" # " # (20)
fd f
= Cx + D u .
gu gd
where L is the duct length. To use this relation in a time domain simulation, a finite-dimensional truncation
is obtained via Padé approximants.
The up- and downstream ends are represented as actuated end elements, as described in Sec. II. In time
domain, Eq. (4) can be written as
" # " #" #
AR 0 BR 0 f
ẋ = x+ , (22)
0 AG 0 BG e
" #
h i h i f
g = CR CG x + DR DG , (23)
e
where matrices with subscripts R and G correspond to state-space realizations of the uncontrolled reflection
coefficient and the actuator transfer function, respectively. Note that R and G for the up- and downstream
ends are different due to the anechoic end mounted at the upstream side.
The pressure at a certain axial location is obtained by inserting a static gain element of the form
fd 1 0 " #
fu
gu = 0 1 (24)
gd
p 1 1
9 of 14
arg(p̂2 /ê)
0.6
|p̂2 /ê| 0
0.4
π
-2
0.2
-π
0
200 300 400 500 600 Hz 800 200 300 400 500 600 Hz 800
frequency frequency
Figure 9. Transfer function from upstream speaker command to pressure in the duct; results from network
model and experiment
Simulation results
The acoustic system was simulated in time domain with reflection coefficients and actuator transfer functions
identified from the frequency response measurements at the experimental facility. Simulations with control
schemes imposing several different acoustic boundary conditions were run.
Figure 10 (left) shows the measured reflection coefficient without control and the results of the simulation
for different prescribed closed-loop reflection coefficients Rcl . Without actuation (black curve) the reflection
coefficient approximately corresponds to that of an open end, i.e., Rol = −1 for low frequencies. The phase of
R corresponds to the length of the duct plus the end-correction of an open end.13 With the control scheme, a
non-reflecting end could be established. For Rcl = 0, the reflection coefficient calculated from the simulation
takes values below 0.05 (red curve), according to a quasi non-reflecting end. (Technically realizable non-
reflecting terminations have reflection coefficients with magnitudes between 0.1 and 0.2.) Discrete K data
and the identified model for the case of tuning the impedance to an anechoic end are shown in the right
frame of Fig. 10. A distinct increase of the magnitude of K for decreasing frequencies can be observed. This
is due to two reasons: i) the uncontrolled reflection coefficient increases for lower frequencies and ii) the
control authority of the speaker that was used deteriorated for smaller frequencies. This could have been
partially overcome by using a woofer.
1.2 10
1 8
0.8 discrete data
6 identified model
|K|
|R|
0.6
uncontrolled
4
0.4 Rcl = 0
Rcl = −1 2
0.2 Rcl = 1, ∆l = 1.3 m
0 0
-π -π
π π
-2 -2
arg K
arg R
0 0
π π
-2 -2
-π -π
250 300 350 400 450 500 550 600 650 Hz 750 200 300 400 500 600 Hz 800
frequency frequency
Figure 10. Right: Reflection coefficients (uncontrolled, no reflection, sound soft, sound hard with additional
virtual length). Left: Controller transfer function to mimic an anechoic end computed from Eq. (7) with
experimental open-loop reflection coefficient and actuator transfer function; discrete data and identified model
are plotted
10 of 14
Anechoic end
Results for the case of an imposed anechoic end with Rcl = 0 are shown in Fig. 11. The experimentally
obtained reflection coefficient magnitudes are slightly higher than the one from the simulation. In the case
of no mean flow, the reflection coefficient magnitude remains below 0.2 in the frequency range considered.
This can be regarded as a non-reflecting end. The reflection coefficient obtained with mean flow (blue curve)
is lower than that without flow, except for frequencies below 350 Hz. Here, the controller was actually only
required to work above 300 Hz. The increase for low frequencies stems from the fact that the loudspeaker’s
response deteriorates (see Fig. 7, right). Also, the controller transfer function has a high gain (Eq. (7)) and,
therefore, the absolute error of the model grows quickly. Better results could certainly be obtained by using
a woofer with a higher response at low frequencies. However, as a result, it can be stated that the impedance
control scheme successfully managed to tune an almost fully reflecting boundary with a pressure node to an
anechoic end.
0.8
0.6
|R|
uncontrolled
0.4 Rcl = 0, M = 0
Rcl = 0, M = 0.05
0.2
0
250 300 350 400 450 500 550 600 650 Hz 750
frequency
Figure 11. Control results for a desired anechoic end. Magnitudes of the uncontrolled and the closed-loop
reflection coefficients with and without mean flow
11 of 14
|R|
0.6 0.6
uncontrolled uncontrolled
0.4 0.4
Rcl = −1, M = 0 Rcl = 1, ∆l = 1.3 m, M = 0
0.2 Rcl = −1, M = 0.05 0.2 Rcl = 1, ∆l = 1.3 m, M = 0.05
0 0
-π -π
π π
-2 -2
arg R
arg R
0 0
π π
-2 -2
-π -π
200 300 400 500 600 Hz 800 200 300 400 500 600 Hz 800
frequency frequency
Figure 12. Boundary condition tuned to fully reflecting, Rcl = −1 (left) and Rcl = +1 (right). In case of the
sound hard end, the phase of the reflection coefficient was additionally tuned so as to mimic an additional
length of 1.3 m
In the left frame of Fig. 12, closed-loop reflection coefficients for an imposed pressure node are presented.
For the case of no mean flow (red curve), the control objective is almost exactly achieved. Over the whole
frequency range considered, the reflection coefficient magnitude deviates less than 5 % from the desired value,
except close to 800 Hz. With mean flow, the results are almost as good as in the no flow case. Overall, the
control scheme manages to establish a pressure node, however, around 260 and 380 Hz, |R| reaches values of
up to 1.2. Less successful results at 380 Hz are probably due to resonance effects in the speaker side channel.
Results for an imposed sound hard end are shown in the right frame of Fig. 12. In addition to a velocity
node, the phase of Rcl was required to mimic a duct elongated by 1.3 m, as in the simulation (see Fig. 10). The
reflection coefficient magnitudes achieved are not as good as in the case of a pressure node. This is reasonable
since the uncontrolled boundary condition more closely resembles a pressure node. Consequently, the control
scheme and the actuator have to perform less when mimicking a sound soft termination. However, good
results are obtained at all frequencies except for the resonance frequency of the laterally mounted speaker
ducts. Also, the phase with the additional delay for an elongated end could be well established. This is
a particularly important case that can be used to generate resonance frequencies which are different from
those imposed by the actual geometry.
0
10
−1
|p̂|
10
uncontrolled
Rcl = −1
−2
Rcl = +1, ∆l = 1.3 m
10
200 300 400 500 600 Hz 800
frequency
Figure 13. Pressure spectra in the duct for the uncontrolled case, an imposed pressure node, and an imposed
velocity node with virtual length of 1.3 m
12 of 14
0
1.2 10
0.8
|R|
α↑ −1
|p̂|
0.6 10
α = 1.0
0.4 α = 0.8
α = 0.6
0.2 α = 0.4
α = 0.2
−2
α = 0.0
0 10
200 300 400 500 600 Hz 800 200 300 400 500 600 Hz 800
frequency frequency
Figure 14. Increasing levels of reflection, realized with a low-reflecting end mounted. Left: Reflection coefficient
magnitude. Right: Sound pressure spectrum in the duct
13 of 14
Acknowledgements
We enjoyed interesting discussions on online wave identification with Gregor Gelbert. Financial support
from the German Science Foundation through the Collaborative Research Center 557 “Control of Complex
Turbulent Shear Flows” as well as from the Bundesministerium für Wirtschaft und Technologie in the
framework of AG TURBO COOREFF-T is gratefully acknowledged.
References
1 Lieuwen, T. C. and Yang, V., editors, Combustion Instabilities in Gas Turbine Engines, Vol. 210 of Progress in Aero-
nautics and Astronautics, AIAA, Inc., 2005.
2 Munjal, M. L., Acoustics of Ducts and Mufflers, Wiley & Sons, Inc., 1987.
3 Poinsot, T. J., Trouve, A. C., Veynante, D. P., Candel, S. M., and Esposito, E. J., “Vortex-driven acoustically coupled
combustion instabilities,” Journal of Fluid Mechanics, Vol. 177, 1987, pp. 265–292.
4 Aurégan, Y. and Starobinski, R., “Determination of Acoustical Energy Dissipation/Production Potentiality from Acous-
tical Transfer Functions of a Multiport,” Acustica, Vol. 85, 1999, pp. 788–792.
5 Guicking, D. and Karcher, K., “Active Impedance Control for One-Dimensional Sound,” Journal of Vibration, Acoustics,
Robust Repetitive Control,” Proceedings of the American Control Conference, Anchorage, AK, 2002.
7 Bothien, M. R., Moeck, J. P., and Paschereit, C. O., “A Modular Approach for Time Domain Modelling of Complex
(Thermo-)Acoustic Systems,” Proceedings of the 13th International Conference on Modelling Fluid Flow, Budapest, Hungary,
2006.
8 Paschereit, C. O., Moeck, J. P., and Bothien, M. R., “State-space Modeling of Thermoacoustic Systems for Stability
Analysis and Time Domain Simulation,” Proceedings of the 13th International Congress on Sound and Vibration, Vienna,
Austria, 2006.
9 Schuermans, B., Bellucci, V., Guethe, F., Meili, F., Flohr, P., and Paschereit, C. O., “A Detailed Analysis of Thermoa-
of Premixed Flames,” Journal of Engineering for Gas Turbines and Power , Vol. 124, No. 2, 2002, pp. 239–272.
11 Gustavsen, B. and Semlyen, A., “Rational Approximation of Frequency Domain Responses by Vector Fitting,” IEEE
ceedings of the 1999 International Conference on Control Applications, MI, USA, 1999.
15 Pota, H. R. and Kelkar, A. G., “Modeling and Control of Acoustic Ducts,” Journal of Vibration and Acoustics, Vol. 123,
Combustors Supported by Active Control,” Proceedings of the 14th International Congress on Sound and Vibration, Cairns,
Australia, 2007.
14 of 14