ACTIVE NOISE CONTROL

SESSIONS
 A New Active Sinusoidal Noise Control System
K. Fujiia and M. Muneyasub
a
Faculty of Engineering, Himeji Institute of Technology, 2167 Shosha, Himeji-shi, 671-2201 Japan
b
Faculty of Engineering, Kansai University, 3-3-35 Yamate-cho, Suita-shi, 564-8680 Japan

This paper proposes a new method to reduce sinusoidal noise components whose frequencies are known. This new method bases
the adjustment on the filtered-x algorithm, therefore requires correctly modeling a replica on the secondary path. However, the
secondary path is continuously changing. The new method hence updates the replica at a specific interval by using the simulta-
neous equation technique that does not require feeding extra noise to the loudspeaker for the update. By slightly loosening the
adjustment, the new method can compose the solvable simultaneous equations on the secondary path. The solutions give the gain
and the phase rotation of the secondary path. This paper finally presents a simulation result to examine the performance of the
new method.

INTRODUCTION f ( j)
Active noise control system reduces primary noise F ( j)
by radiating secondary noise from a loudspeaker. In
this system, the filtered-x algorithm is widely used as a e( j )
means to adjust the amplitude and phase of the secon-
dary noise. A problem is that the filtered-x algorithm a k bk LMS
requires correctly modeling a replica on the secondary
path. In practical system, the secondary path is con- k cos k
tinuously changing. This paper proposes a new method xc ( j )
to be able to reduce sinusoidal noise components under k sin k
the changing. The new method updates the replica at a cos( k jT )
specific interval by using the simultaneous equation k sin k
technique [1, 2] that does not require feeding extra sin( k jT ) xs ( j)
noise to the loudspeaker for the update. This paper fi- k cos k
nally examines the performance of the new method
with computer simulation. FIGURE 1. A configuration of active sinusoidal noise
control systems.
SYSTEM CONFIGURATION
xc ( j ) = (k cosk + k sink ) cos(k jT ) (2)
Figure 1 is a configuration of active sinusoidal noise
and
control systems using the filtered-x LMS (least mean
squares) algorithm. In this system, the primary sinu- xs ( j ) = (k sink + k cosk ) sin(k jT ) , (3)
soidal noise f ( j ) is reduced to e( j ) , by the secon- where k and k are the previously estimated gain

dary sinusoidal noise F ( j ) radiated from the loud- and phase rotation of the secondary path, respectively.
speaker. This paper here assumes that the primary noise A problem is that they are continuously changing.
is expressed as
K SECONDARY PATH ESTIMATION
f ( j ) = k cos( k jT k )
k =1
Figure 2 is an ordinary configuration for estimating
K K the characteristics of the secondary path. A problem in
= a k cos( k jT ) + bk sin( k jT ) , (1) this configuration is that the estimated gains, c k and
k =1 k =1
d k , provide only two equations having four unknowns,
where T is the sampling period, and the radian fre- a k , bk , k and k ,
quency k is known, while the magnitude k and
c k = a k + a k k cos k + bk k sin k (4)
the phase k , i.e., a k = k cos k and bk = k sin k ,
are unknown. d k = bk ak k sin k + bk k cos k . (5)
The filtered-x LMS algorithm updates the gains, a k From the above two equations, the characteristics,
and bk , so as to minimize the error e( j ) by using the k cos k and k sin k , cannot be derived.
filtered reference signals,

SESSIONS
 f ( j) 0

Noise Reduction Effect (dB)

F ( j)
e( j ) -20
(a)

a k bk
-40
cos( k jT ) (b)

-60
ck 0 256 512 768 1024
LMS
dk Block n

sin( k jT ) FIGURE 3. Noise reduction effect provided by the new

method: (a) is the convergence property obtained by using the
characteristics of the secondary path estimated prior to start-
FIGURE 2. An ordinary configuration for estimating the ing the noise reduction. (b) is the noise reduction effect ob-
characteristics of the secondary path. tained by using the characteristics estimated under the noise
reduction.
The simultaneous equations technique [1, 2] states to
( n +1) J ( n +1) J
be able to estimate the characteristics by using the ad- Dn = 10 log 10 e 2 ( j ) f 2 ( j ) (13)
justment error involved in the gains, a k and bk , or j = nJ +1 j = nJ +1
adding small constants to the gains, Four different where J = 320 . In addition, the adaptive algorithm to
gains, a 1k , a k2 , bk1 , and bk2 , obtained by the addition, estimate the characteristics ( c k , d k ) and to update the
give the following equations: gains ( a k , bk ) is both the block Implementation [3] of
c 1 = a + a 1 cos + b 1 sin
k k k k k k (6) k k the filtered-x LMS algorithm whose block length and
step size are 320 samples and 2.0 KJ , respectively.
d = bk a sin k + bk1 k cos k
1
k
1
k k (7)
In Figure 3, (a) is a result obtained by updating a k
c = a k + a k cos k + b k sin k
2
k
2
k
2
k (8) and bk after the characteristics has been estimated
d k2 = bk ak2 k sin k + bk2 k cos k . (9) using the fixed gains a 1k = 0.6 bk1 = 0.6 a k2 = 0.59
and bk2 = 0.59 . In the initial stage prior to starting
From the above equations, we have the solutions:
the active noise control, we can fix the gains to any
(c 1 c k2 )( a 1k a k2 ) + (d k1 d k2 )(bk1 bk2 ) constants like this. Moreover, (b) is a noise reduction
k cos k = k
(a 2 a 1 ) 2 + (b 1 b 2 ) 2
k k k k
effect calculated by using the characteristics of the
(10) secondary path estimated with the gains,
a 1k = a k0 + 0.01a k0 , (14)
(c c )(bk b k ) (d k d k )(a k a k )
1 1 2 2 1 2 1 2
k sin k = k k
, b 1 = b 0 + 0.01b 0 , (15)
(a 2 a 1 ) 2 + (b 1 b 2 ) 2
k k k k
k k k

(11) a = a 0.01a k0
2
k
0
k (16)
where the solutions require the relation: and
(a k2 a 1k ) 2 + (bk1 bk2 ) 2 0 (12) bk2 = bk0 0.01bk0 , (17)
which are fixed adding small constants to a k and bk0
0
SIMULATION RESULT obtained after the convergence shown in (a).
Figure 3 is an example to examine the performance
of the new method using the simultaneous equations REFERENCES
technique to estimate the secondary path. In this exam- 1. Fujii K., Muneyasu M. and Ohga J., Simultaneous
ple, the primary noise consists of fifteen harmonics and equations method not requiring the secondary path fil-
fundamental of 200 Hz, and white noise is added to the ter, in proceedings of Active99, Fort Lauderdale, Flor-
primary noise as disturbance where their power ratio is ida, 1999, pp. 941-948.
30 dB. In addition, the impulse response of the sec- 2. Fujii K. and Ohga J. Method to update the coefficients
of the secondary path filter under active noise control,
ondary path is provided as the sequence consisting of
Signal Processing, 81, no. 2, pp. 381-387 (2001).
32 normal random numbers. 3. Clark G. A., Mitra S. K. and Parker S. R., Block Im-
Under the above conditions, the noise reduction ef- plementation of adaptive digital filter, IEEE Trans.
fect is calculated as Circuits & Syst., CAS-28, no. 6, pp. 584-592, 1981.

SESSIONS
 Discrete-Time Parametric Modeling of Loudspeakers Used for
Active Noise Control in Long Narrow Ducts
Branko Somek, Martin Dadic and Sinisa Fajt

Faculty of Electrical Engineering and Computing, University of Zagreb, Unska 3, HR-10000 Zagreb, Croatia

The secondary path in active noise control systems is generally consisted of electromechanical (loudspeakers) and acoustic
elements. The secondary path is substantial part of any active noise control system. For adaptive systems, it influences on its
convergence properties. For nonadaptive, it makes an additional term in control chain that can cause instability in closed-loop
control. If the duct is long enough, the system will cancel the propagating acoustic wave before it reaches the end. This fact
allows us to use infinite acoustic waveguide model for the duct, and to estimate loudspeaker acoustic load in such a way. The
loudspeaker itself is modeled using electro-mechanical analogies, and obtained continuous-time model is transferred in z-
domain using bilinear transformation. Proposed method can be made to include finite-length waveguides.

INTRODUCTION acoustic load contributed from the reflections from the

duct termination can be ignored [3]. In this case, acoustic
Active noise control (ANC) uses principle of load over the face of the loudspeaker can be determined
destructive interference of the sound waves, e. g. in order from the calculated sound pressure over the square piston
to cancel undesired noise a sound wave with inverse being flush mounted on the wall x=0 wall in the rigid
sound pressure is generated [1]. The low frequency infinite duct of rectangular cross section (0<x<a, 0<y<b)
plane wave sound propagation in ducts makes of them [2]:
the most promising area of application for ANC.

mx
In [2] the theoretical formulas for the sound field in a
hard-walled duct with non-reflecting terminations have
p( x , y , z, t ) = cmn ve jt (2 0m )(2 0n ) cos a

mn
been derived. In [3] a theoretical, frequency domain ny ny ' Ap sin(nd / 2b) 1 jz / cmn 1
model of the loudspeaker in a duct is introduced. cos cos e
b b A (nd / 2b) 2 (d / 2cmn )
In this paper a discrete-time model of dynamic
loudspeakers in long narrow ducts will be developed. e jz / cmn e jd / 2cmn 1 j z / c 1
mn
+ e
2 j 2 (d / 2cmn )

FREQUENCY-DOMAIN MODELING e jz / cmn e jd / 2cmn

(2)
2j
The relationship between terminal voltage and
loudspeaker cone movement can be derived from [3,4]: where the piston is square in shape, of edge length d.
Hence, area of the piston is Ap=d2. Its centre is at (0,y,0)
v( s) s (1) and (x,y,z) are position coordinates (z is axial one) while
H z ( s) = =
E ( s) Lm 3 LD + Rm 2 LK + DR KR a,b are duct cross-section dimensions. Also, ij is
s + s + + Bl s +
Bl Bl Bl Bl
Kronecker delta function, j = 1, is density of air,
where m denotes total effective mass of moving system cmn is mode axial phase speed and is the radian
(i.e. cone, voice coil and effective air mass), D denotes frequency. A is cross-sectional area of the duct.
damping, including mechanical damping in moving Above the natural frequency of the mechanical system
system and radiation resistance, and K denotes effective of the loudspeaker, the cone movement is controlled by
stiffness of suspension and back enclosure, if any. Also L its mass. Although acoustic load on the loudspeaker is
is inductance of the voice coil, l denotes wire length of consisted both from the propagating and evanescent
voice coil, B is flux density in region of voice coil, and v modes, for low frequency range the imaginary part tends
is cone velocity amplitude. R is the resistance of the to be small compared to the mass of the loudspeaker
voice coil, and s denotes complex frequency. cone, and real part tends to be well approximated with
Although perfectly anechoic termination of the duct is Ap2
R MA = c (3)
not realizable, for the reasonably effective attenuator the 2A

SESSIONS
 where c is ambient speed of sound. As the matter of fact, for pulsating sphere[6]. Figure 1. presents both
the acoustic load influences mostly on the cone velocity calculated and measured response. The differences can
in the vicinity of the natural frequency of mechanical be explained with losses in the system and directivity of
system, and for other frequencies very different the microphone.
approximations give similar results.
The propagating pressure contribution for plane waves
is [2,3]:

1 Ap sin( d / 2c))
p( j) = cv ( j) . (4)
2 A d / 2c

The factor sin(d/2c)/(d/2c) can be neglected when the

loudspeaker dimensions are small compared to the
wavelength of sound.

DISCRETE-TIME MODELING

Applying the bilinear transformation [5] a z-domain

transfer function can be derived:

v ( z ) A0 + A1z 1 + A2 z 2 + A3 z 3
H ( z) = = (5)
E ( z) 1 + B1z 1 + B2 z 2 + B3 z 3 FIGURE 1. Frequency response
Lm LD + Rm LK + DR KR
M =8 + 4T + 2T 2 + Bl + T 3
Bl Bl Bl Bl CONCLUSIONS
2T 2
2 T 2
A0 = A1 = , A2 = A3 = A discrete-time parametric model of loudspeaker flush
M M
LD + Rm 2 LK + DR
Lm KR mounted on a wall in the rigid duct is developed. The
24 4T + 2T + Bl + 3T 3 loudspeakers are substantial part of secondary and
Bl Bl Bl Bl
B1 = feedback paths in ANC, and such a model allows time-
M domain simulation of ANC systems. It also can be useful
Lm LD + Rm LK + DR KR
24 4T 2T 2 + Bl + 3T 3 in some approaches to the ANC systems synthesis.
Bl Bl Bl Bl
B2 =
M REFERENCES:
Lm LD + Rm LK + DR KR
8 + 4T 2T 2 + Bl + T 3
Bl Bl Bl Bl
B3 = 1. Nelson P.A. and Elliott S.J., Active Control of Sound ,Academic
M Press, London,1992, pp. 116-160.

2. Doak P. E., Excitation, Transmission and Radiation of Sound from

where T is sampling period. Source Distributions in Hard Walled Ducts of Finite Length (1):
Although it might seem to be advantageous to include Effects of Duct Cross-section Geometry and Source Distribution
acoustic load on the loudspeaker in a more accurate way, Space-Time Pattern, Journal of Sound and Vibration, 31,1-72 (1973)
the differences are very small (below 1db), and , more 3. Shepherd I.C., Cabelli, A. and LaFontaine R.F., Characteristics of
over, approximation of the loudspeaker with the square Loudspeakers Operating in an Active Noise Attenuator, Journal of
piston is not accurate itself. Sound and Vibration, 110, 471-481 (1986)
Since perfectly anechoic termination of the duct is not
4. Locanthi B.N., Application of Electric Circuit Analogies to
realizable, the exact experimental verification of the LoudspeakerDesign Problems, Journal of the Audio Engineering
model is not possible. Beside numerical examinations, Society, 19, 778-785 (1971)
we made some experiments with a rigid finite 70 cm long
5. Oppenheim A.V. and Schafer R.W., Digital Signal Processing,
duct with flush mounted broadband loudspeaker. The Prentice-Hall International (UK), London, 1975, pp. 206-211.
both ends of the duct were opened, and acoustic load on
the loudspeaker and sound pressures in the duct were 6. Crandall I. B., Theory of Vibrating Systems and Sound, D. van
Nostrand Company, Inc., Princeton, 1954, pp. 95-103, 149-152.
calculated using transmission lines theory [6] and
loudspeaker and duct cross-sectional area ratio, similar
as it is used for the reflex enclosure loudspeakers [4].
The radiation impedance on the ends was estimated as

SESSIONS
 A Study of Active Noise Control in Ducts using Digital
Signal Processing
R. M. Velloso Camposa , E. Bauzer Medeirosb
a
Departamento de Engenharia Eletrnica PUC-MINAS - Av. Dom Jos Gaspar 500, 30535-610, Belo
Horizonte, MG, BRAZIL, tel: + 55 31 33194115, fax: + 55 31 33194224, email: cvr@mandic.com.br
b
Departamento de Engenharia Mecnica da UFMG- Av. Antnio Carlos 6627, 31270-901, Belo Horizonte, MG,
BRAZIL, tel: + 55 31 34995247, fax: + 55 31 34433783, email: flugzbau@dedalus.lcc.ufmg.br
The present work describes some of the research effort on Active Noise Control (ANC) being jointly developed by the
Catholic University of Minas Gerais (PUC- MINAS) and the Federal University of Minas Gerais (UFMG). Considerations
about the implementation of Digital Signal Processing for noise control in ducts has been presented. The objective is to
establish a study on Active Noise Control in ducts combining geometry and acoustic parameters modification together with
adaptive digital filtering implementation. The main results are presented and considered according to their use in developing
real applications. The idea is to provide an initial and useful insight for both designers and students concerned about Active
Noise Control in ducts. The authors are continuing with this research, for other configurations and implementations and believe
that the present studies should hopefully provide more favourable conditions to the implementation of real time A N C.

INTRODUCTION extremity of the duct, near the acoustic ending, which

was varied during the test. Figure 1 shows a closed end
In recent years Active Noise Control (ANC) has termination. The microphones and the loudspeakers
become a feasible reality for a variety of practical were connected to a Texas Instruments DSP-TMS
applications. The development of reasonably priced 32C6211 board connected to a PC, providing the
powerful processors and efficient software has enabled required signal processing.
implementation of real time ANC [1].
ANC in ducts has deservedly earned a place of CONTROL PROCEDURE
importance for a variety of reasons. Sound carrying
ducts occur in a great number of important applications Adaptive filtering as initially proposed in the
and the existing boundaries (the duct walls) may lead classical work of Wiener [2] has been the chosen
to simplifications in the wave propagation studies. solution. Figure 2 indicates a block diagram containing
However ANC in ducts still presents some difficulties, the main elements of the control system.
particularly when wide band noise is being analysed. In
contrast to narrow band noise, wide band noise may
occur in certain applications and is considerably more U n fi lter ed Estima- E s t. S i gnal
ted
difficult to cope with, demanding efficient signal Noi se System
processing.

EXPERIMENT OVERVIEW Noise

Error
Noise LMS
Filter S i gna l
Testing was carried out in a set up represented in
Figure 1, where one of the geometrical configurations
used can be seen. System
Re al
Fi lte red No i se

F I G U R E 2 : C o n t r o l S ys t e m D i a g r a m
FIGURE 1: Test set up
Noise at the input has been generated by means of a
white noise generator. The white noise output was
The primary source, a loudspeaker, was positioned
filtered before being input into the system, enabling
at the left duct inlet, as shown and the control source,
appropriate choice of the input signal frequency range,
another loudspeaker at the inlet of the T joint. A
which was kept within the lower frequency spectrum,
microphone was positioned near the end of the right
for well known reasons [2].

SESSIONS
 Input signals for the adaptive FXLMS ( Filtered-X Th e d u c t c h a r a c t e r i s t i c s h a v e a l s o b e e n
Least Mean Square) algorithm were obtained from the evaluated, producing results which can be seen
data acquisition procedure. A FIR (Finite Impulse in Figure 5.
Response) digital filter was employed with the upper
cut-off frequency set at 600 Hz. A single microphone
was employed in the process, to provide the error
signal. Plant identification compared with the (also)
known noise input signal enabled the generation of the
secondary source (control) signal, which was then
input at the second loudspeaker.
The adaptive LMS algorithm may be conveniently
described by equation (1), where was set to 0.014, FIGURE 5: Duct Identification
which was found to be appropriate for this experiment
[3],[4]. Once the plant had been characterized it was
w(m+1) = w(m) + (y(m)e(m)) (1) possible to generate the noise control signal (Figure 6),
which was then input into the noise control source.
Implementation of the LMS algorithm has been
carried out, according to the following sequence:
! Initial parameter choice
! Adaptive filter output evaluation
! Error signal evaluation
! Adaptive algorithm evaluates weighing
! System identification
! Noise cancelling signal generation
Equation 2 has been used to establish adaptive filter
output.
L 1
(2) F I G U R E 6 : N o i s e C o n t r o l S i gn a l
y ( n ) =
lk = 0
w k (n ) x ( n l)
Th e p r e vi o u s l y s h o wn F i gu r e s i n d i c a t e a
t yp i c a l s e t o f r e s u l t s o b t a i n e d fr o m t h e t e s t s e t
u p o f F i gu r e 1 . O t h e r t e s t s h a v e a l s o b e e n
RESULTS AND CONCLUSIONS c a r r i e d o u t , wi t h a m e a s u r e o f s u c c e s s , fo r o t h e r
ge o m e t r i c a n d s i gn a l c o n d i t i o n s .
Using the previously described procedure the system Th e p r e vi o u s l y d e s c r i b e d p r o c e d u r e i s n o w
was conveniently described by Figure 3, which was b e i n g v e r i f i e d a n d r e fi n e d i n o r d e r t o i mp r o v e
previously known, at the beginning of the test. noise cancellation response, including the
d e t e r mi n a t i o n o f o p t i m a l s e n s o r a n d s e c o n d a r y
s o u r c e p o s i t i o n i n g fo r a va r i e t y o f g e o m e t r i c
c o n fi gu r a t i o n s .

ACKNOWLEDGEMENTS
The authors would like to thank PUC-MG for the
support given to the present research. They are also
indebted to Rodrigo Costa Ivo, and Alexandre Augusto
Moraes Nogueira for their vital participation in the
F I G U R E 3 : P r e v i o u s l y K n o wn S y s t e m
development of the present research
System identification produced Figure 4, which
confirmed the previously assumed system REFERENCES
characteristics and, consequently, a confirmation of the
1. C. H. Hansen, S.D. Snyder, Active Control of Noise and
algorithm efficiency.
Vibration, E & FN Spon, 1997.
2. N. Wiener, Interpolation, Extrapolation and Smoothing
of Stationary Time Series, John Wiley & Sons, 1949.
3. S.V. Vaseghi, Advanced Signal Processing and Digital
Noise Reduction, John Wiley & Sons, 1996.
4. KUO, Sen M & Morgan Dennis R., 1996, "Active Noise
Control System Algorithm and DSP implementations",
John Wiley & Sons.
FIGURE 4: System Identification Results

SESSIONS
 Active Control of Double-glazing Windows: an
Experimental Comparison between Feedback and
Feedforward Controllers
Andre Jakob, Michael Moser
Institut f
ur Technische Akustik, Technische Universit
at Berlin, Einsteinufer 25, 10587 Berlin, Germany
andre.jakob@tu-berlin.de

Double-glazing windows are distinguished by high acoustic transmission loss for the high- and mid-frequency range
but weak in the low-frequency range especially around the mass-spring-mass resonance frequency of the panel-cavity-
panel system. In the work presented the cavity sound field of a double-glazing window is actively minimized by
means of secondary loudspeakers and error microphones inside the cavity. The influence of the minimization on the
transmission loss of the double-glazing window was investigated experimentally. In this presentation will be reported
the experiments performed with multichannel feedback control as well as those with multichannel feedforward control.
A comparison will be given. In the case of the feedforward controllers the reference signal was obtained either by
means of a microphone in the sending room of the window testing facility or directly from the signal generator. This
leads to difficulties due to the properties of the sending room which will be discussed.

INTRODUCTION
In recent years some authors have investigated
the active control of double-glazing windows by
means of loudspeakers inside the cavity between the
two panels (cf. [1, 2, 3, 4]). In this paper a compar-
ison will be given between the use of two adaptive
feedforward controllers and one adaptive feedback
controller. The difference between the two feedfor-
ward controllers is the generation of the reference
signal, which was obtained either directly from the
signal generator as in [1, 3] or from a microphone in
the sending room of the window testing facility. In
all three cases a controller with 4 loudspeakers and
4 error microphones was used. The positions of the
loudspeakers and microphones were near the corners
of the window. All used filters were FIR filters and
were adapted using the well known multiple error
LMS algorithm.

FIGURE 1: Test setup with feedforward controller.

FEEDFORWARD CONTROL
In Fig. 1 is depicted the experimental setup con-
sisting of the double-glazing window built into the troller hardware used allowed a maximal length of
window testing facility and the adaptive feedfor- the (four) adaptive FIR filters of 256 coefficients.
ward controller. Six microphones in the receiving The results of the tests, measured in third-octave
room were used to measure the mean sound pres- bands, are shown in Fig. 2. The maximum level
sure level with and without active control measures without control is found at 100Hz, because in this
applied. The primary excitation was band limited third-octave band the mass-spring-mass resonance
white noise. frequency of the plate-cavity-plate system is in-
For a first test the reference signal of the feedfor- cluded. In this region the sound insulation of the
ward controller was taken directly from the signal double-glazing window is comparatively low. The
generator. This setup is a somewhat idealized ar- idealized feedforward control yielded a high level
rangement and in general will yield some kind of reduction in all measured third-octave bands. The
best case results. In the second test the reference reduction of the total sound pressure level was
signal is taken from an additional microphone in 9.5dB.
the sending room in front of the window. This is Upon employing a reference microphone, the im-
a more realistic case. In both test cases the con- provements in the 100Hz third-octave band were

SESSIONS
 FIGURE 3: Test setup with feedback controller.
FIGURE 2: Measured sound pressure level in the
receiving room with and without active control. Exci-
tation with band limited white noise. computational complexity compared with the feed-
forward controller only 128 filter coefficients could
be realized with the existing hardware. The re-
nearly as high as in the case before but in the upper duction of the number of filter coefficients nega-
frequency range the improvements were much lower. tively influences the performance of the controller
In the 160Hz third-octave band was found even a (cf. Fig. 2). The reduction of the total sound pres-
higher sound pressure level than without control. sure level was measured to only 3.3dB.
The reduction of the total sound pressure level was Nevertheless, some tests were performed with
only 4.7dB; the reasons being, firstly, that there is a traffic noise, e.g. from highways, trains, jet aircrafts
feedback from the loudspeakers inside the cavity to and helicopters, using the adaptive feedback con-
the reference microphone that was not compensated troller (see also [4]). As always is the case in adap-
for, and secondly, that the position of the reference tive active noise control, the achievable performance
microphone certainly is in one of the many nodes of highly depends on the nature of the signals. Espe-
the sending room modes. Out of these, the latter is cially for traffic noise signals, which consist of high
the most important, which could be verified through harmonic components, reductions of the total sound
the transfer function between primary and reference pressure level of more than 8dB were obtained with
signals, showing a minimum around 160Hz. The re- the adaptive feedback controller.
sults show that the relative position of the reference
microphone with respect to the sending room field ACKNOWLEDGEMENTS
heavily influences the achievable level reduction by
the active system. In a practical application how- This work was sponsored by the DFG, Deutsche
ever the reference microphone certainly would be Forschungsgemeinschaft, contract No. Mo390/6-1.
built into the outer side of the window frame at a
fixed position. This means that this position would
mainly determine the measured transmission loss
and of course would result in completely different
transmission losses in other testing facilities. Free REFERENCES
field conditions should be approximated in the send-
ing room, in order to yield comparable results, but [1] P. De Fonseca, W. Dehandschutter, P. Sas, and
this is generally not the case in window testing fa- H. Van Brussel. Implementation of an active noise con-
trol system in a double-glazing window. In ISMA 21,
cilities. pages 377388, 1996.

FEEDBACK CONTROL [2] S. Pietrzko and O. Kaiser. Experiments on active control

of air-borne sound transmission through a double wall
cavity. In Active 99, pages 355362, 1999.
The problems when using a reference microphone
can be overcome employing a feedback controller [3] A. Jakob and M. M oser. Enhancement of the transmis-
instead of a feedforward. Here the adaptive feed- sion loss of double panels by means of actively controlling
the cavity sound field. In Active 99, pages 363374, 1999.
back controller given in [5] has been investigated.
This controller uses the error signals and the com- [4] A. Jakob and M. M oser. Active control of the cavity
pensation signals for an estimate of the disturbance, sound field of double panels with a feedback controller.
which in turn serves as the reference signal for the In 7th International Congress on Sound and Vibration
ICSV7, 2000.
adaptive algorithm. Fig. 3 illustrates the setup with
the feedback controller. The block called refer- [5] S.M. Kuo and D.R. Morgan. Active noise control sys-
ence signal estimator includes an additional filter- tems. John Wiley & Sons, 1996.
ing with all the secondary paths. Due to the higher

SESSIONS
 An Improved Method for Off-line Secondary and Feedback
Paths Estimation for Active Noise Control System
*Hui LAN Ming ZHANG Wee SER

Centre for Signal Processing

S2-B4-05, Nanyang Technological University, Singapore 639798
*ehlan@ntu.edu.sg

In the Active Noise Control (ANC) system, the transfer function of feedback and secondary paths are always
identified in off-line before adaptation of the ANC controller. In [1], an off-line method is presented to identify
both feedback and secondary path simultaneously by inserting an additional white random noise to secondary
loudspeaker. Two adaptive filters are used to identify both paths. However in case that it is impossible to turn off
the primary noise, the adaptation of the two adaptive filters will be affected by the primary noise, especially the
one to identify feedback path, since reference microphone is very close to the noise source. Therefore it is highly
desirable to reduce the affect of primary noise to path identification. When the primary noise is periodic, its
influence can be reduced by an adaptive line enhancer [2]. In this paper, two adaptive line enhancers are applied
to remove the influence of primary noise to identification of both feedback and secondary path. Moreover, to save
the computation power, two filters are not updated every sample. Simulation results show that more accurate
transfer functions for both paths can be obtained, especially for the feedback path.

and A/D converter. The feedback path begins at input

INTRODUCTION of secondary loudspeaker and consists of D/A
converter, power amplifier, secondary loudspeaker,
Active Noise Control (ANC) system attempts to acoustic path from secondary loudspeaker to reference
reduce acoustic noise by introducing a second sound microphone, pre-amplifier, reference microphone and
source, which emits sound with same amplitude but A/D converter. The existence of secondary and
inverse phase to the original sound. Feedforward ANC feedback paths causes potential instability of ANC
system is the most popular. A single input single output system, so they have to be compensated. Although
feedforward ANC system has two microphones and a feedback path can be removed by using some physical
secondary loudspeaker. A reference microphone senses methods, such as special arrangement of secondary
the noise to be controlled, called primary noise, and loudspeaker or using directional reference microphone,
provides prior knowledge about the primary noise to etc. [1], sometimes the reference microphone
the controller. An error microphone measures the unavoidably affected by the anti-noise from secondary
supposition result of primary and secondary noise, and loudspeaker. On the other hand, the secondary path
also provides the adjustment information to the always appears.
controller. To be adaptive to the changing environment,
adaptive filters are always used as ANC controller. PROPOSED METHOD
Based on the information provided by reference and
error microphones, the ANC controller adjusts its To compensate the effect of secondary path, the
weights and filters the reference signal to produce a filtered-x LMS (FxLMS) algorithm and its modified
proper signal so as to drive secondary loudspeaker. The version are proposed [3][4][5][6]. The basic idea is to
loudspeaker emits an anti-noise signal, which destructs filter the reference signal x(n) by a estimated model of
primary noise around the position of error microphone. secondary path S(z), S ( z ) , and use the filtered
The most popular adaptive algorithm used in ANC reference signal to update the weights of ANC
system is the Least Mean Square (LMS) algorithm due controller, W(z) by the LMS algorithm. A way to
to its simplicity and robustness. When the LMS compensate the effect of is feedback path neutralization
algorithm is applied as the ANC controller, there exist [1], where the effects of feedback path F(z) is
two electro-acoustic paths, i.e., secondary path and neutralized by a estimated model of this feedback path,
feedback path. Secondary path starts from input of F ( z ) .
secondary loudspeaker to output of error microphone.
The model of secondary and feedback path, i.e.,
It includes D/A converter, power amplifier, secondary
loudspeaker, acoustic path from secondary loudspeaker S ( z) and F (z ) can be estimated by using adaptive
to error microphone, pre-amplifier, error microphone filters without operation of the ANC controller W(z).

SESSIONS
This is called offline estimation of S ( z ) and F ( z ) . A another one is for reducing influence of d(n) to
way to identify both paths at the same time by using secondary path modeling, which is expressed as H2(z).
adaptive filters is described in [1]. A white random Two delay units with taps Df > L and Ds > M are used
noise v(n) is used as excitation to modeling adaptive to decorrelate the signals due to v(n).
filters S ( z ) and F ( z ) . The desired signal for secondary
REFERENCES
path e(n) is obtained by error microphone, while that of
feedback path x(n) is sensed by reference microphone.
When primary noise does not exists, S ( z ) and F ( z ) can [1] S.M. Kuo and D.R. Morgan, Active Noise Control
Systems -- Algorithms and DSP Implementations,
model S(z) and F(z) quickly. However when primary New York: Wiley, 1996.
noise exists, it becomes an interference to adaptation of [2] Kuo, S.M.; Vijayan, D., ``A secondary path
both S ( z ) and F ( z ) . In case the primary noise is modeling technique for active noise control
periodic, it can be removed by using adaptive line systems,'' IEEE Trans. on Speech and Audio
enhancer [5]. This idea is proposed in [2] for online Processing, vol. 5, No. 4, pp. 374 -377, July 1997
secondary path modeling. In this paper we extend it to [3] B. Widrow and S.D. Stearns, Adaptive Signal
offline secondary and feedback path modeling. Processing, Englewood Cliffs, NJ: Prentice-Hall,
The proposed algorithm using line enhancer to 1985.
reduce influence of narrowband primary noise to [4] S. J. Elliott, Signal Processing for Active Control,
offline identification of secondary and feedback path is New York: Academic Press, 2001
illustrated in Figure 1. Two adaptive line enhancers are [5] S. Haykin, Adaptive Filter Theory, 3rd ed, Upper
applied. One is for removing effects of x(n) to feedback Saddle River, NJ: Prentice-Hall 1996
path modeling, which is denoted as H1(z), while
d(n)
P(z)
u(n)

vf(n) vs(n)
F(z) S(z)

+ +
+
+ + +
u(n) e(n)
x(n) x(n-Df) e(n-Ds)
z-Df H1(z) H2(z) z-Ds
+ +
- -
+ +
xc1(n) ec2(n)
xe(n) LMS LMS
ee(n)

+ +
vfh(n) vsh(n) -
+ -
(z)
F S (z) +
vef(n)
ves(n)
LMS LMS

v(n)

Figure 1. Block diagram of proposed method

SESSIONS
 The Problem of Choosing Sensor Configuration
in Active Sound Control Systems :
a Simple Numerical Illustration

V. Martin
Acoustics and Electromagnetism Laboratory (LEMA)
Swiss Federal Institute of Technology, CH-1015 Lausanne, Switzerland

It has been shown recently that different sensor configurations can be made equally efficient in controlling an acoustic domain
by weighting the primary pressure. There exists an optimal weighting which maximizes the robustness while still maintaining the
efficiency. The very simple numerical application proposed allows us to verify some assertions that could seem surprising.

made up of the secondary source responses to the Nc

INTRODUCTION c
control microphones and p n the reference primary
pressures at these sensors, and let D be the weighting
Recent works have shown that all sensor matrix that modifies the pressures. What D leads to An
configurations of an active sound control system could in W from an LMS control with Nc sensors ?
be made efficient by weighting the primary pressures, A first method consists in reducing to zero the
and also that robustness against spatial variations of the 2
functional J (p ) = G .n + D .p
c c

primary field could be maximized [1]. To illustrate res n c

when n=fn n

these results, a very simple arbitrary numerical leading to D .p n = G c .H .G .p n . An infinity of full

c -1 *

application is proposed, along with additional

information obtained. D satisfy the equality while only one diagonal does.
2
c
Minimizing J res (p n ) = G c .n + D .p n
c
at fn also
SENSOR EFFICIENCY
leads to a solution. In this case,
THROUGH WEIGHTING c
Gc . D .pn = H c .H .G .p n . An infinity of matrices D ,
-1 *

In auto-adaptive active sound attenuation systems, diagonal or full, can satisfy the equality.
the secondary source driving signal(s), or control, is Having chosen a possible weighting D, the optimal
achieved by minimizing the total pressure at the attenuation of a primary field in W, obtained from Nc
control microphones. Ideally a large number N of sensors, is expressed by
them are located inside W where attenuation is sought. *
p .( I - A + A c ).p
With G the responses of the secondary source to the N A c (p) = - 10 log10 *
where
sensors, and I the identity matrix, the control fn that p .p
-1 * -1 * * -1 -1 * -1 *
best minimizes the primary field pn, called here of Ac (D) = (Hc .Gc .D.P - H .G ) .H .(Hc .Gc .D.P - H .G )
reference, solves the following problem: with D such that A c (p n ) = A n .
2
min G.f + I .p n = min J res (p n ) Let us consider a very simple arbitrary situation. The
f f

Whatever p , the optimal attenuation has the form domain W is made up of 3 microphones. The reference
* primary pressures and the secondary source responses
p .( I - A ).p
opt
A (p) = - 10 log10 are respectively 4.,2.,6. and 1.,2.,3. The configuration
*
p .p with both sensors M1 and M3 is chosen. Dealing with
where A=G*.H-1.G and H=G*.G. It arises from the diagonal matrices D only, we obtain
opt -1 *
optimal control f = - H . G .p . In particular, for pn, 13 0.

s 0
we obtain A (p n ) = A n .

D = 28 13 and D = 0 65 - 2s
opt

In reality, only a small number Nc of sensors is 0. 63 9

available (Nc <N). Where should they be placed in 14
order to achieve An when dealing with pn? Let Gc be

SESSIONS
 Let us consider the configuration M1-M2 with D
Thus, J res ( p n ), J res (p n ) and J res (p n ) , the latter for an
65.
arbitrary value of s, reach their minimum at fn . diagonal. Efficiency at pn is assured by v = ( - s)
28.

MAXIMIZING ROBUSTNESS and N (D ) - A subsequently has the form

Let us work p = pn + dp with . Of all values of e 1.

opt
leading to the same value of dA , one is minimum
s 1. 13. 2. 2. s 3.
2. - , ( - )- , - .
and called emin. Figure 1 shows all the possible values 5. 14. 2. * 7. 14. 5. 14.
opt
of A ( p ) when all variations dp are admissible. The 3.
s 1. 13. 2. 2. s 3.
minimum attenuation A (e min ) decreases when emin
min
thus min ( - + ( - )- + ).
increases. s 5. 14. 2. * 7. 14. 5. 14.
op t 1.9643
0.
Using Mathematica, D = . By
0.
0.4167
exhaustive calculation, curves A c (e min ) via D
min

opt
diagonal and via D are obtained. Figure 2 shows
A c (e min ) , increased by optimization. Configuration
min

M1-M3 already robust with D diagonal has not been

FIGURE 1. Set of optimal attenuations
improved, while configuration M2-M3 has reached
(e min )
min
with lower boundary A
A (e min ) .
min

The minimum attenuation is the solution of

FIGURE 2. Robustness maximization
A min
(e min ) = min A (p )
opt
for configuration M1-M2
p E

dp
E = { p = p n + dp , ( ) min e min}
p
With only a small number of sensors to control W,
the minimum attenuation in W, A c (e min ) , still
min

decreases when emin increases, but now there exists a

maximum value of emin beyond which we cannot
guarantee the absence of amplification. A c (e min )
min CONCLUSION
solves the same type of problem as above. Finally, additional information has been obtained :
A sensor configuration is all the more robust that it
guarantees a better minimum attenuation. Maximizing there exists a weak form D of the weighting matrix
the robustness consists thus in finding an optimal which makes it possible, when diagonal, to maximize
weighting matrix D leading to the highest curve robustness as defined in the text. In the particular
example chosen for the illustration, the solution can
A c (e min ) . At best, it is similar to A (e min ) , i.e., Dopt
min min

almost be obtained by hand.

*
results in Ac = ( N - A ) ( N - A ) zero with N(D) . In
REFERENCE
other words S S Nij ( D ) - A ij = 0. . Numerical
i j

experiments have shown that solving the problem 1. Martin V., and Gronier C, Sensor configuration
efficiency and robustness against spatial error in the primary
min S S N ij (D) - Aij = 0.
D i j field for active sound control ,
to be published in Journal of Sound and Vibration, 2001

with D = D or D
leads to a Dopt which maximizes robustness, letting
us understand that the smaller the norm N(D) - A ,
the greater the robustness.

SESSIONS
 A generalised Model of Sound Emission Transducers in
active Noise Control and Arrays

V. Adam, O. Meylan and M. Rossi

Laboratory of Electromagnetics and Acoustics (LEMA)

Swiss Federal Institute of Technology (EPFL), CH-1015 Lausanne, Switzerland

This paper deals with a generalised model of electroacoustical sound emission transducers for their design and measurements
in the fields of ANC and arrays. This model, able to describe the behaviour of a transducer both in reception and emission
modes, is based on the representation of the electromechanical (eg. electrodynamic, electromagnetic, etc.) and
mechanoacoustical couplings by two-port networks, with associated two-terminal elements. In active noise control a major
difficulty is to take into account the effects of the primary noise on the transducers, particularly regarding phase relationship.
The same difficulty arises in loudspeaker arrays. A practical problem is to design the transducers in such a way that these
effects are reduced to a minimum. The generalised model allows these problems to be solved by representing the primary field
using a real source eg. an ideal pressure source with an associated impedance instead of modifying the radiation impedance
as is usually done. Simplified diagrams, obtained by applying network theory, proved to be powerful and useful tools for the
design and characterisation of sound actuators.

BASIC THEORY given in the figure 4, where Fg and Zmg represent the
transducer. In ANC, the need for improved modelling
Most of electroacoustic sound sources are made of of transducers is recognised, mainly in view of their
an electromechanical transducer and a design (definition of their requirements) and testing,
mechanoacoustical coupling involving a radiating face. taking into account the effects of external sound fields:
The main assumptions for their study are: kd < 1, k, a transducer can be subjected to the noise to be
wave number, d, characteristic dimension (eg radius of counteracted as well as many other neighbouring
a circular piston), quasi linear domain (small non- transducers. If we consider the resulting external sound
linearities). In the following, we use symbols, pressure pe, a net force Fe is exerted on every
quantities, units, conventions and signs according to transducer, which can always be expressed as Spe.
the IEC. Usually, as seen in the literature, the effect of external
fields is taken into account by modifying the
They can be described by lumped-constant circuits transducer radiation impedance, i.e. through the
with two-port networks which represent the impedance Zml. That means the melting of two
electromechanical and mechanoacoustical couplings different sources of forces: the reaction to the radiation
[1]. Equivalent circuits can be drawn up that only (that is the fundamental meaning and raison dtre of
include elements and quantities of the same nature by the radiation impedance) and the external action. We
removing the coupling networks. Finally, very simple think that this is not the best way of modelling and that
circuits are obtained using circuit theory. Figure 1 it is preferable to maintain the dissociation.
gives the basic circuit of a sound emission transducer.
A two-port network connects an electrical system, DESIGN
represented by a real source (Ug, Zg), and an
mechanical/acoustical system or medium, represented Figure 5 shows how these effects can be simply
by a mechanical impedance Zml equivalent to the represented, as previously, by a load impedance and a
acoustical load. For example the acoustical radiation force source Fe. With ad hoc values of the elements,
impedance becomes Zml = S2Zar , with an appropriate design can be carried out taking into account the
value of the radiating face equivalent area S. Figures 2 external field effect. For example, if sound pressure
and 3 give the networks which correspond to the two cancellation is required on the radiating face, the
fundamental linearized couplings, that is: a) force transducer is short-circuited: (vve) is maximised and
controlled by the current, - electrodynamic, the displacement too. In consequence the effects of
electromagnetic and magnetostrictive conversions; b) non-linearities are augmented. The ratio Zme/Zml and
force controlled by the electric charge, - electrostatic the value of Fg determine the ability of a transducer to
and piezoelectric conversions. These circuits can be tackle this extreme condition. For example for an
reduced to a simpler one by Thevenins theorem, as electrodynamic transducer, the most severe

SESSIONS
perturbation occurs around the resonance frequency. source and a near-field microphone measures the
These effects are minimised by current control instead sound pressure very close to the transducer. By acting
of voltage control [2]. on magnitude and phase of the transducer excitation,
the minimum sound pressure level (or any pertinent
v
Zg I value) at the microphone position is reached.
Two-p o r t
Ug U Netwo r k F Zml This test has been proposed within the framework
TPN of a BE European project and proved to be successful
for transducer assessment. It was used to characterise
some limitations of prototypes when interacting
FIGURE 1 Basic transducer circuit strongly with the an external field, in this case with
almost no front acoustic pressure. Different transducer
I Ze Zm v technologies (direct radiator, compliant structures, air-
n
flow modulation, electrodynamic, piezoelectric) have
been assessed using this set-up.
U F

Figure 6 presents typical spectra which are

obtained with a transducer under such a test. The
fundamental frequency has been reduced by 20dB,
FIGURE 2 TPN for force driven by current while the harmonic content increases. As has been
theoretically described, this measurement shows

I Ze
n
Cm Zm v clearly that in the presence of an external sound field
the transducers exhibit increasing non-linear behaviour
Co F by getting closer to their working condition limits.
U

150

FIGURE 3 TPN for force driven by charge

v Zmg Zml

SPL
(dB)
Fg

FIGURE 4 Simplified final circuit 50

300 1k 10k

Frequency (Hz)

(v - ve) Zmg Zml

FIGURE 6 Near-field SPL measurement with (bold)
and without (fine) an external sound field (1900Hz
Fg Fe sine wave).
AKNOWLEDGMENTS

FIGURE 5 Circuit including external effects The work presented here has been supported by the
Swiss Federal Education and Science Office (OFES),
to whom we are grateful.

TESTING REFERENCES
1. Rossi M., Acoustics and Electroacoustics, Norwood:
To test the ability of a sound emission transducer to Artech House, 1988, pp. 309-353.
be used in a ANC system, a measurement set-up was
built. The transducer under test is mounted on a baffle 2. Adam V., Comportement dun haut-parleur soumis un
with an auxiliary source generating an external sound champ extrieur in Proceedings of the 5th French
field. A sine wave signal is applied to the auxiliary Congress on Acoustics, Lausanne: edited by PPUR,
2000, pp. 660-663.

SESSIONS
 A Practical, Fast and Cost-efficient Algorithm for Multiple
Input, Multiple Output Active Noise Control Applications
P. Sjstena , S. Johanssonb , S. Nordebob and I. Claessonb
a The National Institute for Working Life, P.O. Box 8850, 40272 Goteborg, Sweden
b Blekinge Institute of Technology, P.O. Box 520, 372 25 Ronneby, Sweden

Active noise control has proven to be an efficient solution to low frequency noise problems in many different applications. A large part
of these applications are concerned with harmonic noise control and typically require control systems with several inputs and outputs.
As the systems grow, the demand for processor capacity increases rapidly, resulting in large and expensive hardware platforms. As
multiple-input, multiple-output noise control is restricted to the control of periodic noise, a controller structure that is adapted to
the signal type may significantly reduce the requirements on the hardware capacity. This paper discusses a complex, time-domain
controller that is designed for the control of harmonic components. The structure of the controller is simple, easily implemented and
can easily be extended to handle any number of noise references and any number of harmonics.
The convergence properties of a multiple-input, multiple-output control system depends largely on the acoustic coupling between the
active sources and the control sensors. To get accurate and stable control it is necessary to use normalization, i.e. a weighting function
that optimizes the controller for each control source. With the presented controller structure, the use of normalization is straight-
forward and a number of different approaches for normalization is discussed as well as examples from practical inplementations.

INTRODUCTION THE COMPLEX LMS CONTROLLER

The foundation for the discussion in this paper is a Traditionally, one composite reference signal is gen-
feed-forward, adaptive controller. Such controller struc- erated, containing all harmonics to be controlled. The
tures have been extensively examined and discussed in reference signal is fed through an FIR filter with suffi-
the literature since the birth of the LMS algorithm in the cient number of filter weights to adjust the magnitude and
late fifties. A feed-forward controller implies that there phase for each harmonic.
is a reference signal that is correlated to the disturbance. In the proposed controller structure, each harmonic is
The control output is applied to the environment through considered to be a separate reference signal and is treated
an actuator. The influence on the plant by the actuator equivalently to references from different sources[1]. The
is measured by one or more control sensors. The discus- reference generator produces one complex reference sig-
sion in this paper is directed towards the use of several nal for each harmonic and each reference is controlled
control outputs and several control inputs (MIMO). The by one complex weight. Thus, the (real) signals at the L
controller structure can also handle reference inputs from outputs y(n) = [y1 (n) yL (n)]T are given by
any number of noise sources. R

The standard LMS algorithm is derived for random in- y(n) = {xr (n)wr (n)} , (1)
r=1
put signals. Since MIMO control implies the control of
harmonic components, it makes sense to adapt the con- where xr is the complex reference signal for reference r,
trol scheme to the signal type. Periodic reference signals wr (n) = [wr1 (n) wrL (n)]T is the L 1 vector of com-
can be generated internally within the controller. There plex weights for this reference and R is the number of
are huge advantages with this method compared to mea- references. {} denotes the real part of the complex
suring the reference signal with some kind of sensor: The multiplication which implicates that in a practical imple-
reference signal will contain only those harmonics that mentation only the real part is evaluated. One may note
are to be controlled and the properties of these signals, that the output signal is simply obtained as the sum of the
i.e. frequency and signal power, are known. The refer- contributions from all references.
ence signal is generated by using a synchronization sig- The LMS update algorithm is derived in the same
nal that is taken from the noise generating mechanism, manner as for the real LMS algorithm. For reference r
e.g. a rotating machinery. The synchronization signal can this results in
be used to control the sampling rate for the controller (or- wr (n + 1) = wr (n) 2Mr xr (n)F H
r e(n) (2)
der based control) or to generate reference signals with
the appropriate frequencies in a controller with a fixed where F r is a M L matrix of complex gain elements,
sampling rate. describing the change in magnitude and phase between

SESSIONS
each output and every control sensor input, for the fre- real elements, given by
quency that relates to reference r. The hat implicates that
the elements in F are estimated (i.e. measured) values. 0000
rl = . (6)
e(n) = [e1 (n) eM (n)]T is the M 1 vector of real con- r M 2
m=1 |Frml |
trol sensor inputs and Mr is a square (L L) convergence These elements are calculated off-line and require only
factor matrix. The formulation of this matrix is further one (real) multiplication for each output controller (and
discussed below. The leakage factor is close to one reference).
and is very useful in practical implementations, since it In a practical application, the sum of the squares of
restrains the behavior of the weight vector w[6]. the complex gains, as given by equation (6), may well be
larger than the sum of the cross terms, i.e. complex gains
between different outputs and control inputs. In this case,
NORMALIZATION the matrix F H F is diagonally dominant, which leads to[3]

The convergence characteristic for a multiple input, diag F rH F r F H
r Fr . (7)
multiple output controller is determined by the acoustic
(or structural-acoustic) conditions, i.e. the complex gain Under these circumstances, the algorithm obtained by us-
between each control output and every control sensor in- ing the convergence matrix given by (5) has a perfor-
put. Variations in the complex gain between outputs re- mance that is comparable to the Newton algorithm.
sults in different convergence rates for different outputs,
which in turn effects the overall attenuation[3][4].
In general, the quantity that governs the convergence PRACTICAL APPLICATIONS
rate and the stability
of the controller
is given by the Hes-
sian matrix E xr (n)F rH F r xr (n) [2]. Under the given as- The suggested normalization leads to a very simple
sumptions, the reference signal xr (n) is deterministic and algorithm with no matrix operations and with a mini-
for the ordinary MIMO LMS controller this expression mum of complex operations. It can easily be extended
reduces to[6] to any number of inputs, outputs and references (har-
monics). The algorithm has been implemented i C on
Mr0 = 00 (r trace F rH F r )1 I (3) a TMS320C32 floating point signal processor for use in
e.g. a silent seat application and to reduce noise from the
where I is an L L identity matrix, 00 is a conveniently rotating cutter in lawn mowers. The algorithm has also
chosen positive convergence factor, and rh denotes the been used to control propeller noise in a mock-up of a
power in the reference signal (which is known). Equation twin-prop aircraft[4].
(3) ensures a stable but slow controller, since the update
algorithm is given the same normalization for all outputs.
Another alternative is the Newton-like REFERENCES
algorithm[2][6], which is obtained when the convergence
matrix is chosen as 1. B. Widrow, J. M. McCool and M. Ball, The Complex LMS
Algorithm, Proc. IEEE 63, 719-720 (1975)
M00r = 000 (r F H 1
r Fr ) . (4)
2. P. A. Nelson, S. J. Elliot, Active Control of Sound. London:
By combining equations (2) and (4), it is obvious that Academic Press, Inc., 1992.
the full matrix of complex gains is used in the update 3. S. Johansson, Active Noise Control in Aircraft, Algorithms
of each output controller. This is a very powerful algo- and Applications, Licentiate thesis, Department of Signal
rithm, but the computational cost is fairly high since a Processing, University of Karlskrona/Ronneby, 1998.
full complex matrix multiplication must be performed at 4. S. Johansson, P. Sjsten, S. Nordebo and I. Claesson, Com-
each sampling interval. parison of Multiple- and Single-Reference MIMO Active
In the third alternative presented here, the update al- Control Approaces using Data Measured in a Dornier 328
gorithm for a specific output is normalized with the com- Aircraft, Int. J. Acoust. Vib. 52, 77-88, (2000).
plex gains that are related to that particular output[4]. The 5. S. Johansson, I. Claesson, S. Nordebo and P. Sjsten,
convergence factor matrix becomes Evaluation of Multiple Reference Active Noise Control Al-
gorithms on Dornier 328 Aircraft Data, IEEE trans. on
Mr000 = 0000 (r diag F H 1
r Fr ) (5) Speech and Audio Proc., 74, 473-477, (1999).
H 6. B. Widrow, S. D. Stearns, Adaptive Signal Processing,
where the matrix diag F F is the diagonal matrix con- PrenticeHall, Inc., 1985.
sisting of the diagonal elements of F H F. This matrix has

SESSIONS
 Fundamental Consideration of Active Noise Control
Based on Division of Input Signal
Noboru Nakasako and Masayuki Tatebe
School of Biology-Oriented Sc. & Tech., Kinki Univ.
Nishi-Mitani 930, Uchita-cho, Naga-gun, Wakayama 649-6493 Japan.
The Filtered-x LMS algorithm is very often adopted for the active noise control, but it requires the estimation of transfer function of
error path C in advance. In this study, differing from the Filtered-x LMS algorithm, a new method for ANC based on division of input
signal is proposed, i.e. algorithm which identify the objective unknown system h and transfer function of error path C simultaneously
by dividing input signal into two components. The input vector is divided into two components by introducing an adequate coefficient
matrix and signal vector. One component is used for estimating h and the other for C. Finally, the effectiveness of the proposed method
is confirmed through the simulations.

     
        
INTRODUCTION with

     
xk xk
 
1 xk N 1

   
Recently, the ANC system[1] which controls an ob- xk 1 xk 2 xk N

      
XN N k ..
jective noise mainly in a low frequency region by sound .
becomes very popular in many engineering fields accord- xk N 1 xk N xk 2N 2
ing to the rapid development of the technique in Signal

 (2)

Processing, especially DSP (Digital Signal Processor). d k

Source
The Filtered-x adaptive algorithm is generally and very K z

yk ek
often adopted to update a single-channel Feed-forward xk
ANC system on controlling the noise in a duct, etc. The Processor C z
Filtered-x algorithm usually needs to estimate in advance


Adaptive

a system model of error (secondary) path (so-called C z C z Algorithm

filter) between the loudspeaker and the error microphone,
because the system C z is considered as a priori infor-


mation. However, it is inevitable that this C z filter has
FIGURE 1. Adaptive ANC system.
hk

modeling error due to the temporal change of transfer By applying the LMS method, the successive estima-
function, etc.[2, 3] Therefore, the ANC system deterio-
rates its performance and sometime becomes unstable in
tion algorithm can be derived. However, since this LMS



algorithm does not utilize directly the input noise x k but
the filtered input XN N k CN , the coefficient vector CN


the worst case.
In this study, differing from the usual Filtered-x LMS must be estimated as C N.
algorithm, a new method for ANC based on division of Now, by considering an arbitrary signal as the com-
input signal is proposed, i.e. algorithm which identify the position of random vectors[4], it can be represented in a
objective unknown system h and transfer function of error linear combination of noise vn as follows:
path C simultaneously by dividing input signal into two
xk
 

K
an k vn (3)

components. The input vector is divided into two compo- n 0
nents by introducing an adequate coefficient matrix and Here, an k denotes a coefficient and K is the number of

signal vector. One component is used for estimating h data points in each frame. We consider the input signal
and the other for C. Furthermore, the effectiveness of the x k as the arbitrary signal. Equation (3) can be expressed



 !"


proposed method is confirmed through the simulations. in vector-matrix notation as follows:
DIVISION OF INPUT SIGNAL xk a 1 k v1 a 2 k v2 aK k vK Va k (4)

Let consider the ANC system as shown in Fig.1. Here,




K z denotes the transfer function from the detection sen- Then, the coefficient vector a k can be calculated
sor to the error sensor and C z from the loudspeaker to (though is unknown) as follows:

 #






the error sensor. a k V 1x k (5)
Let x k be the input and d k the desired output, the
error e k is given by





When the random vector matrix V is assumed to be the

sum of two random vector matrices V V1 V2 , x1 k  

ek
 d k 
 

d k H zC zxk
XN N k CN T
hN (1) x1 k

and x2 k can be set respectively as:
V1 a k x2 k V2 a k %
$


(6)

SESSIONS
On estimating h and C, the effective information on the

.
Next, at the 2 107-th iteration, the transfer function


input can be divided into x1 and x2 , though the error signal
e k at the error sensor is entirely same as that in Eq.(1).
This means that x2 is an ambient noise when we regard
C z of error path is changed. That is, we increase all the
coefficients of impulse response for C z with 10 % and

give their sign at random. Figure 3 shows the estimation
x1 as the effective information, while x1 is ambient when
accuracy of parameters and the noise reduction. After the
regarding x2 as the effective information.
parameter change, we can see that both EA and Reduction
Thus, the recurrence estimation algorithms for the un-
are recovered up to the levels before changing.
known system h and the filter C can be obtained respec-



'&
( )


tively as follows: 300

EA (dB)

'&
( )


1
h N k 1 h N k e k X N k 200



kC (7) NN
2 100
N k
C 1 N k
C e k XN N k h N k (8)

( )


( )

0
Here, is a constant related to the convergence speed. 0 1 2 3 4
1 2 Iteration 7
The matrices XN N k and XN N k correspond to the di- [10 ]
vided inputs x1 k and x2 k , and the error signal can be

Reduction (dB)
300


 

constructed as follows: 200
e k d k XN N k CN T h N (9) 100
0
-100
DIGITAL SIMULATION 0 1 2 3 4
Iteration 7
[10 ]
300
FIGURE 3. EA of h (Top) and Reduction (Bottom).
EA (dB)

200
100 Finally, the estimation procedure is also demonstrated
0 when the S/N (Signal to Noise Ratio) is set to 20 (dB) as
0 1 2 3 4 shown in Fig.4. It is found that the proposed method can
Iteration 7
[10 ] also reduce the objective noise.
Reduction (dB)

300 30
200
EA (dB)

100 20
0 10
-100 0
0 1 2 3 4
Iteration 7
[10 ] 0 1 2
Iteration 7
[10 ]
FIGURE 2. EA of h (Top) and Reduction (Bottom). 100
Reduction (dB)

50
The effectiveness of the proposed method is confirmed
through digital simulation. The tap lengths of the sys- 0

  *

tems are set as 128 and the number of data sample is 4

+

(K 3). As input xk , the time series with AR model
-50
0 1
Iteration 7
2

 ,
xk 1 bx k k of 1st order is adopted (where [10 ]
b 0 5 and is the Gaussian random number with 0 FIGURE 4. EA of h (Top) and Reduction (Bottom).
mean and variance 1). To evaluate the performance of
algorithm, the estimation accuracy(EA) and the noise re-
duction(Reduction) are introduced as: REFERENCES

 # - # 
$
 
N 1 N 1
h2i 2 1. P.A. Nelson and S.J. Eliott, Active Control of Sound, Aca-
EA 10 log10 hi h i (10)

 - 
,
demic Press, London 1992.
i 0 i 0
2. B. Widrow and S.D. Stearns, Adaptive Signal Processing,
Reduction 10 log10 dk2 dk yk 2
(11)
Prentice Hall, New Jersey 1985.

 ,. #
Figure 2 shows the estimation accuracy of parameters
and the noise reduction ( 0 1 10 4). In this case, it
is obvious that the good results in both EA and Reduction
3. S. D.Snyder and C.H. Hansen, IEEE Trans. Signal Process-
ing 42, 950-953 (1994).
4. A. Tokura, O. Mitura, and M. Tohyama, Proc. Acoust. Soc.
are obtained. Jpn. 1, 497-498 (1999.9) (in Japanese).

SESSIONS
 Identification of Secondary and Feedback Paths in ANC
O. Jircek, M. Brothnek
Department of Physics, CTU-FEE in Prague, Technick 2, 16227 Prague, Czech Republic
The paper deals with the identification and realization of secondary and feedback paths in multi-channel active noise control systems.
A feed-forward ANC system with microphones as reference and error sensors requires identification of the transfer function from
the digital controller outputs to corresponding inputs to ensure convergence of the algorithm used; a filtered-x LMS algorithm was
assumed. Three different off-line methods of identification were tested and compared: identification using white noise, identification
using MLS signals and training by adaptive algorithm. The algorithms and arrangements were tested on an experimental duct of
rectangular cross-section of dimensions similar to actual ventilating or air-conditioning ducts. The source of noise was realized by
means of an axial flow fan with five blades, with the flow speed being regulated by regulation of fan revolutions. The tested algorithms
were implemented on a digital control unit based on Texas Instrument floating point DSP.

secondary error
INTRODUCTION sources microphones
reference testing
fan microphone
microphone
One of the first applications of active noise control
(ANC) is the attenuation of fan noise in a duct. If the
cross-section of the duct is large enough, higher modes x y2 e1 e2
y1
have to be assumed in the frequency range of ANC, and
a multi-channel system has to be assumed. If the feed- DSP

forward ANC system is used, a precise identification of

analyzer
all error and feedback paths is of crucial importance for
the system performance. The off-line identification and
FIGURE 1. Arrangement of the two-channel ANC system.
modelling is reated in this paper.

the measured response of the chosen path and the re-

IDENTIFICATION AND MODELLING
sponse of the model filter.
The second posssibility of identification is based
Off-line identification can be performed by three dif-
on measurement of impulse response by MLS signals.
ferent methods: identification using white noise gener-
MLS signals produced by the controller (DSP unit) pass
ated by a noise generator, identification using MLS sig-
through the chosen error of feed-back path and then the
nals produced by the controller, and identification based
impulse response is calculated. First N points of mea-
on training by adaptive algorithm. Identification process
sured impulse response correspond to N coefficients of
results in impulse responses or transfer functions of error
FIR filter if measured on the same sampling frequency.
and feed-back paths respectively. They can be realized
The length of the impulse response of FIR filter is
by FIR of IIR filters in the controller. FIR filters are al-
determined by the number of coefficients N. However,
ways stable and have a linear phase, but the lenght of the
impulse responses of both error and feed-back paths can
impulse response is determined by the order of the fil-
be very long, because the transfer functions are affected
ter. Difficulty with the stability and design of IIR filters
by the reflections on the input end of the duct. To ob-
is compensated by a longer useful impulse response for a
tain good approximation of mentioned transfer functions,
lower number of coefficients.
very long FIR filters are necessary.

FIR filters modelling IIR filter modelling

FIR models of error and feed-back paths can be ac- In the two-channel ANC system depicted in figure 2,
quired by means of an adaptive training algorithm in six model filters must be used. Sufficiently long (large)
which the white noise is sent to the secondary loud- FIR filters spend significant portion of controller mem-
speaker and picked up by the appropriate microphone. ory and computational time. This can be solved by us-
The adaptive algorithm minimizes the difference between ing IIR filters, which can be stable only if their poles lie

SESSIONS
 0.06
measured by MLSSA
within the unit circle. The IIR filter coeficients can be ob- IIR model

0.04
tained by solution of Prony equations to minimize a error
function. As these equations are nonlinear with respect to 0.02

both nominator and denominator coefficients the standard 0

Prony method is used to obtain suboptimal solution. For -0.02

precision of the filter coefficents, the Steiglitz-McBride

-0.04

method itarative method is used. In practical implemen-

tation, the filter poles can exceed the unit circle due to -0.06
50 100 150
n
200 250 300

the approximation (rounding) error. This efeict has to be FIGURE 4. Measured and modelled feed-back path F1 .
considered in DSP realization. An example of error path
poles is shown in figure 3. Comparison of measured and
modelled feed-back path can be seen in figure 4.
rate up to 1.5 cubic meters per second was used. Flow
speed was regulated by regulation of fan revolutions. A
F1
high pass filters of 100 Hz were implemented at all inputs
y1 (n ) of the controller [2]. For modelling of error paths, the IIR
G1 filters of 20th order were used. For feedback paths, the
order of 40 was chosen. An example of attenuation of fan
C11 e1 (n ) noise is shown in figure 5. In this article modelling of
LMS e2 (n ) error and feedback paths using IIR filters for two-channel
x (n ) C12 ANC in a duct was presented. Attenuation results ob-

tained show the system to perform well, achieving atten-
C21 e1 (n ) uation up to 15 dB.
LMS e2 (n )
C22 100
control-off
control-on

90

G2 y2 (n )
80
L[dB]

70

F2 60

50

FIGURE 2. Block diagram of two-channel ANC algorithm.

40
0 100 200 300 400 500 600 700 800

f[Hz]

FIGURE 5. The noise reduction at flow speed of 7 ms 1.

ACKNOWLEDGMENTS
This work has been partially supported by the research
program of the CTU Prague J04/98:212300016 and par-
tially by the GACR research project No.102/01/1370
Multichannel systemof active noise control.

z
REFERENCES
FIGURE 3. Poles of error path model filter.
1. Nelson, P., A., Elliott, S., J.: Active control of sound, Aca-
demic Press, New York, 1993.
2. Ai, X., Qiu, C., Hansen. C., H.: Minimizing wind effects on
EXPERIMENT active control systems for attenuating outdoor transformer
noise, Noise Control Eng. J., 48(4), 2000.
Experiments were performed on a duct of rectangular 3. Jircek, O., Brothnek, M.: Error and feed-back paths mod-
cross-section of dimension of 200 300 mm. An axial eling using IIR `lters in ANC system , Proc. of POSTER
flow fan with five blades and maximum volumetric flow 2001, CTU-Prague, 2001.

SESSIONS