4th European Conference on Noise Control

EURONOISE 2001
Patra, 14-17 January 20001
c EURONOISE

ACTIVE NOISE CONTROL BASED ON FUZZY MODELS

C.A. Silva, J.M. Sousa and J.M.G. Sá da Costa

Technical University of Lisbon, Instituto Superior Técnico

Department of Mechanical Engineering/ GCAR
Av. Rovisco Pais, 1049-001 Lisboa, Portugal
Tel: +351 21 8419119, Fax: +351 21 8498097
E-mail: fcs,j.sousa,[email protected]

Key words:Active noise control, fuzzy model-based control, fuzzy clustering.

Abstract. This paper presents a new approach to acoustic noise control, by introducing a fuzzy
model-based control strategy. Classical linear identification and control tools have been ap-
plied to active noise control in the last two decades. In this type of control, the limitations of
their applicability are well defined. Therefore, new techniques must be developed in order to
increase the performance of active noise controllers. Recent research in control has emphasized
the importance of nonlinear model-based control, increasing the performance of several types
of systems. From the different nonlinear techniques, fuzzy modeling is one of the most appeal-
ing. Fuzzy clustering can be used to identify models directly from data. These models proved to
be accurate for complex and partly known systems, and can represent highly nonlinear systems
in an effective way due to their function approximation properties. This paper applied fuzzy
and classical modeling to real-time noise data. The fuzzy model revealed to be more accurate
than the linear model. An inverse model is derived which is applied in the proposed model-
based control scheme. The paper presents control results derived from an experimental setup,
revealing best performance than classical control methods.

1
1 INTRODUCTION
Technological and industrial development has resulted in an increase of noise level from ma-
chines, factories, traffic, etc. Many reasons have been taken into account to develop some
research in the sound attenuation area, such as environmental and legal restrictions concerning
people’s safety and health.
There are two ways to achieve noise reduction. The first one are the passive systems, which
are based on the absorption and/or reflection properties of materials, and present excellent re-
sults for frequencies higher than 1 kHz. The other consists of active systems, which show good
performance for lower frequencies and are based on the principle of interference, where a sound
with the same amplitude as the noise source but an opposite phase is generated in order to cancel
the primary noise. This is usually known as Active Noise Control (ANC).
Most of the ANC applications use linear identification and control techniques, relying on the
linear superposition of a primary field and a destructively interfering secondary field to achieve
the desired noise cancellation. In practical terms, this means that one pressure fluctuation will
not become distorted by the presence of another. However, in some situations it is not possible
to neglect the nonlinear terms in the wave equation [6]. Elliott [5] describes two types of
nonlinearities present in a system: Weak nonlinearities, where the behavior of the system under
control is mainly linear but can have nonlinear components, e.g., an harmonic distortion due
to saturation or hysteresis of the actuator. In these cases the principle of superposition fails
when pressure fluctuations of very large amplitudes and propagation over very large distances
are taken into account. These nonlinearities can be solved using a compensation term in the
control loop; Strong nonlinearities which are driven by tonal excitations producing responses
which waveform has a continuous broadband frequency spectrum.
This paper deals only with weak nonlinearities. Therefore, fuzzy models will be adopted as
dynamic acoustic models, since they are specially suited for nonlinear mapping, and are able to
deal with nonlinear effects such as distortion and reverberation.
In order to investigate possible developments in active noise control of enclosures, this paper
presents the simulation of noise reduction in a square wood box, using a control technique based
on a fuzzy model.
The outline of this paper is as follows. In section 2 a short review of classical active noise
control is presented. Sections 3 presents a brief overview on fuzzy modeling. Section 4 de-
scribes the application of fuzzy models to active noise control. The experimental setup is pre-
sented in section 5 and section 6 shows the experimental results. Finally, in section 7 some
conclusions are drawn and future applications of nonlinear identification techniques in ANC
are discussed.

2 CLASSICAL ACTIVE NOISE CONTROL

The earliest ideas for active noise control are outline in the patent granted to Paul Lueg in 1936
[8]. But it was only in the 80’s, when the digital controllers became accessible for everyone ,
that the active noise control became efficient.

2
 ANC strategies can be divided into two groups: feedforward systems, where a reference
signal is available from a detection sensor and the information of a monitoring sensor is used to
adapt the controller, and feedback systems, where there is no detection of the noise source. The
signals from the sensors are fed back to the actuator which in turn produces an acoustic signal
at the sensor which is added to the signal produced by the noise source [8].
In the present-day ANC applications, the controllers are usually Finite Impulse Response
(FIR) adaptive filters, where a feedforward control loop is used. These filters are moving aver-
age models of the actuators and are designed to minimize a cost function, usually the acoustic
pressure measured by a mesh of sensor microphones [8]. This strategy can be easily generalized
for the multivariable case when the cancellation problem requires several speakers and micro-
phones. When a feedback loop is used, the filters are designed in the frequency domain using,
for instance, robust control analysis [12].
ANC was already implemented successfully in several systems, such as air conditioning
ducts to attenuate the low frequency noise produced by fans, or in transport systems, such as
cars or planes, to reduce the sound produced by the engines. Both the fan and engine noises
are harmonic sounds, which make the noise reduction a relatively simple problem. However,
in many other applications the sound attenuation would be desirable but the sound field is far
more complex, such as in enclosures or free field applications. The classical controllers have
their bounds of application well defined and the use of linear time invariant model based control
does not allow for further improvements in terms of control performance. Nonlinear modeling
techniques may bring a new insight and develop the possibilities of active noise control.

2.1 Single channel feedforward control

In single channel feedforward control it is assumed that the signal used to drive the actuator
is derived from a detection sensor. This detection sensor captures an electric signal, which
provides a prediction of the noise at the control point. This detection signal can be measured
by a microphone located upstream the controller in the propagating sound field, a mechanical
transducer on a rotating machine, an accelerometer, an optical sensor, etc. At the control point
it is also introduced an microphone, which is called the error sensor. This error signal is nor-
mally used to monitor the performance of the active control system or to adapt the controller
parameters.
The idea behind this control strategy is very simple. In the usual feedforward loop, the
controller is fed by the reference signal and generates a control action to the plant. Usually the
()
controller used in this case is a FIR filter w n , which is a function of a discrete instant n, and
is defined by
X
I i

w(n) = a
f (n
i i) (1)
i=0

()
where ai are the filters coefficients, f n is the signal input of the filter, and I is the total number
of filter coefficients.

3
 d(n)

x(n) + e(n)
u(n)
w(n) G(z) S
+

r'(n)
G'(z) X
Figure 1: Feedforward loop control with LMS adpative controller

In order to obtain better active controllers, the error signal can be used to adapt the FIR
controller, using the least mean squares algorithm. The adaptation of the filter is derived by
filtering the detection signal using a model of the system.
()
The error e n to be minimized is given by the pressure at the control point, which has two
() ()
terms: the primary noise pressure d n and an estimate of the output y n :
e(n) = d(n) + y (n) (2)
The estimate of the output uses the reference signal r (n) and the adaptive version of the FIR
filter to be used on-line wi :
X
I i

y (n) = w
r (n
i i) : (3)
i=0
() ()
The reference signal r n is the detection signal x n defined previously filtered by a model of
the actuator gbi :
X
I i

r(n) = gb
x(n
i i) : (4)
i =0
Figure 1 shows the described control loop. In fact, this type of controller acts like an inverse
model controller [8]. Note that in spite of using the error signal, the control loop is said to be
feedforward, since the error is not directly used by the controller.

3 FUZZY MODELING
3.1 Fuzzy logic – basic principles
0
Fuzzy logic is a generalization of yes-no Boolean logic. Assigning to false values and to 1
true ones, fuzzy logic also allows in-between values. Assuming that U is a set of values of
u, fuzzy logic defines a mapping from U to the unit interval through a membership function
( ): [0 1]
A u U ! ; . This membership function can be defined in linguistic terms. As it can be
seen in Fig. 2, the subjective feeling of noise as a result of a sound pressure level of 120 dB is
not defined in a crisp sense as Noisy or as an Intolerable sound - it is defined as : Noisy and : 05 05
Intolerable. In this way, the uncertainty present in every systems can be implicitly represented.

4
 y

Low Medium High

u
1

0
m Bi

m Ai

0 70 90 140 u

Quite Noisy Intorelable

Figure 2: Linguistic fuzzy model.

3.2 Fuzzy models

Fuzzy models are flexible mathematical structures that, in analogy to neural networks and radial
basis functions, have been recognized as universal function approximators [14]. Fuzzy models
use ’If-Then’ rules and logical connectives to establish relations between the variables defined
for the model of the system. For the given example, let the system to model be the relation
between the noise sensation and the sound pressure level. Thus, in fuzzy modeling the fuzzy
’If-Then’ rules take the form:
If u is Noisy then y is High (5)
The fuzzy sets in the rules serve as an interface amongst qualitative variables in the model,
and the input and output numerical variables. The fuzzy modeling approach has several ad-
vantages when compared to other nonlinear modeling techniques, such as neural networks: in
general, fuzzy models can provide a more transparent representation of the system under study,
maintaining a high degree of accuracy.

3.3 Takagi-Sugeno (TS) fuzzy model

Takagi and Sugeno [13] introduced a fuzzy rule-based model that can approximate a large num-
ber of nonlinear systems. The Takagi-Sugeno (TS) fuzzy model consists of representing the
base rules as follows:
R i : If u is A i then y = f (u); i = 1; : : : ; K
i (6)
where Ri denotes the ith rule, K is the number of rules, u is the antecedent variable, y is
the consequent variable and Ai is the antecedent fuzzy set of the ith rule. Each rule i has a

5
 y

x+b3
y=a 3

b
x+
2

y=a 2

b 1
x+
a1
y=
u

m Ai

0 70 90 140 u

Quite Noisy Intorelable

Figure 3: Fuzzy TS model with clustering

different function fi yielding a different value for the output yi . The most simple and widely
used function is the affine linear form:

R i : If u is A i then y = a u + b ; i = 1; : : : ; K
T
i i (7)

where ai is a parameter vector and bi is a scalar offset. Figure 3 presents an example of a TS

model.

3.4 Model identification using fuzzy clustering

A fuzzy model can be constructed entirely on the basis of systems measurements. In the fol-
lowing, we consider data-driven modeling based on fuzzy clustering [1, 2]. Further, this model
can be simplified in order to decrease computation time in control applications [9].
Assume that data from an unknown system is observed. The aim is to use this data to
construct a deterministic function to approximate the system . The model of the system can be
described by a Nonlinear Auto Regressive with eXogenous input (NARX) model of the form
given in (8)

y^(k + 1) = f (u(k n ); : : : ; u(k

d n ); y (k); : : : ; y (k
u n y )): (8)

where nd ; nu and ny represent the number of pure delays, delays at the input and delays at the
output respectively.
()
The time signals of the output y t and the input u t are sampled at a constant sampling ()
rate. The identification problem consists of predicting how changes in u k affect y k , given () ()
by the function f .

6
 Clustering consists of grouping data into clusters. Generally, a cluster is a group of data ob-
jects, which are more similar to one another than to other data in different clusters. This fuzzy
similarity should be understood as mathematical similarity, measured in some well-defined
sense, for example by using a distance norm. Then, fuzzy clustering consists of assigning a
rule-based function between the input data and the output data to describe a NARX model as in
(8). This function can be parameterized by a Takagi-Sugeno (TS) model as in (7).
Assuming that the input and the output variables are known, the nonlinear identification
problem is solved in two steps:

 Structure identification - consists of transforming the dynamic identification problem into

a static nonlinear regression, by choosing the systems structure: input delays, output
delays and pure delays.

 Parameter estimation - is done by fuzzy clustering, where data is divided in K clus-

ters using the Gustafson-Kessel algorithm [3]. The antecedent membership functions are
extracted by projecting the clusters in the respective variables, and the consequent param-
eters in (7) are obtained by local or global mean-squares [1].

Figure 3 presents a possible clustering of the data used to identify a TS fuzzy model.

4 FUZZY MODEL-BASED ACTIVE NOISE CONTROL

The classic feedforward active control is based on the model of the actuator, a reference signal
and an error signal. In section 2 it was stated that feedforward control loop may be seen as an
inverse model controller. In order to apply this control strategy an accurate inverse model must
be available. Recently, some techniques to derive inverse fuzzy (nonlinear) models have been
proposed [2, 7]. The use of nonlinear models improves in general the accuracy of the model
of the plants under control. Since the derivation of the NARX model is by no means restricted
by stability or invertibility conditions, inverse nonlinear models can be easily derived by using
fuzzy clustering.
Nonlinear acoustic models of the process, including reverberation, can be identified using
nonlinear modeling techniques such as fuzzy modeling. With this acoustic model allows the
controller to predict the noise at the control point. The model of the plant allows also the
application of an internal model control strategy. This type of strategy was already proposed
by [4], using artificial neural networks. Fuzzy clustering was already successfully used in [10]
to obtain very accurate models of acoustic actuators. A different acoustic control approach is
proposed in [11], where fuzzy models are used to attenuate sound using a predictive control
strategy.
The control loop proposed in this paper is depicted in Fig. 4, where both the acoustic model
(^ ) ()
A n of the acoustic path and the inverse model of the plant G 1 n are identified using fuzzy
modeling, see Section 3.

7
 Acoustic plant
d(n)

x(n) ^ + e(n)
d(n) u(n)
^
A(n) G -1(n) G(n) S
+

Figure 4: Inverse model controller.

5 EXPERIMENTAL SETUP
5.1 The enclosure
In order to perform accurate measurements and test active noise control algorithms, an ex-
perimental setup is required. For this purpose, a closed and isolated box with a volume of
09 07 04
:  :  : m3 was developed. This box must receive the less possible sound from any other
source but the loudspeakers. Its construction guarantees perfect isolation from the outside. The
isolation is achieved by using heavy or dense materials. Two wood walls with different thick-
ness (19 and 12 mm) were used, which have also an air interface between the walls. To avoid
undesirable reflections, the air interface was filled with an absorbing material (glass wool with
25 mm of thickness). The sound field produced consists of sound (noise) at one position inside

Figure 5: The closed isolated box.

the box using a loudspeaker. Another loudspeaker at a different position produces the control
actions that will cancel the noise at particular points inside the box. In order to apply this con-
trol strategy, a model of the loudspeaker is required. The model represents the output of the
loudspeaker at a microphone placed in a specified point inside the box, where it is desired to
cancel the sound when an input voltage is applied.

8
5.2 Loudspeakers and microphones
To implement active noise controllers in transport systems, one of the most strong requirements
is that the price of the control system is low - the devices must be as cheap as possible, but nev-
ertheless robust. The chosen speakers are typical dynamic loudspeakers and the microphones
are of the electret type. The experiment used 1 loudspeaker to control, one to play as noise
source and one microphone. The microphone was placed 0.33 m away from both noise and
control loudspeakers.

5.3 The measurement chain

The experimental data was acquired using a National Instruments NI6024-E data acquisition
board and the Matlab’s toolbox Data Acqusition. A Wavetek Rockland spectrum analyzer was
used to generate white noise. The simulation and control programs were developed in Matlab,
where the fuzzy clustering toolbox developed by [1] is utilized to derive the nonlinear fuzzy
models. All the experiments were done with a samplig rate of 2 kHz.

6 EXPERIMENTAL RESULTS
6.1 Modeling of the plant
The excitation signal used to derive the models was white noise. The accuracy of the models
is measured using the root mean square error (RMS), and the percentile variance accounted
for (VAF). The VAF measure rates the variance  of the error between two signals, the original
signal y and the modeled signal y m , and the variance of the original signal y . Both measures
were done using the power spectrum density of the signals S : , since the analysis must be done
in the frequency domain. Therefore, in this paper the VAF is defined by
  (S S m )

VAF = 1 y

 (S )
y
 100% : (9)
y

Note that when the real output power spectrum S y and the models output power spectrum S ym
are equal the VAF has the value of 100% . This measure is adimensional, contrary to the RMS.
The experimental results are presented for the bandwidth [100,1000] Hz, because the used loud-
speakers have low gain for lower frequencies.
A linear FIR filter model and a TS fuzzy model are derived using the same identification
data. The FIR filter has I = 128 coefficients and the TS fuzzy model has the following orders:
nu = 40 = 40
; ny and nd =2 . The pure delay value is chosen based on the plant physics; the
sound speed has a standard value of 331.5 ms 1 [6].
Since the frequency rate is 2 kHz, the sound takes 2 samples to travel the distance between
the loudspeakers and the microphone (0.33 m, see section 5). The modeling results are shown
if Fig. 6 and in Table 1. The TS fuzzy model has a RMS value lower than the FIR model, but
a slightly lower VAF. However, Fig. 6 shows clearly that the TS fuzzy model copes better the

9
 −40

Power Spectrum Magnitudde [dB]

−50

−60

−70

−80

−90
200 400 600 800 1000
Frequency [Hz]

Figure 6: White noise modeling. Continuous – real output; dashed – FIR model; dot-dashed –
TS fuzzy model

FIR filter model TS fuzzymodel

RMS (dB) 2,72 2.06
VAF 86,7% 86.0%
Table 1: Model accuracy measures.

higher frequencies. Moreover, the orders of the fuzzy model are much smaller, which can be an
advantage in terms of implementation.

6.2 Identification of acoustic and inverse models

The control approach proposed in section 4 needs the acoustic path model and the plant inverse
model. The acoustic plant model is a TS fuzzy model which has the same orders of the model
of the plant: nu= 40, ny and nd = 40 =2
. Note that a FIR model could also be identified here.

The inverse model is identified directly from data using the direct inverse learning as de-
scribed in [7]. This model is a TS fuzzy model with the following orders: n u , ny = 80 = 80
and nd =0 . Figure 7 presents the acoustic path model and Fig. 8 the inverse model. Table 2
presents the models’ accuracy measures.
The direct model derived in the previous section and the acoustic model have a similar
accuracy. The inverse model presents an excellent result, except for 200 Hz. This implies that
this control strategy can only be used for noise sources with frequencies higher than 200 Hz.

10
 −30

Power Spectrum Magnitudde [dB]

−40

−50

−60

−70

−80

−90
200 400 600 800 1000
Frequency [Hz]

Figure 7: Acoustic path modeling. Continuous – real output; dashed – TS fuzzy model

−10

−20
Power Spectrum Magnitudde [dB]

−30

−40

−50

−60

−70

−80

−90
200 400 600 800 1000
Frequency [Hz]

Figure 8: Inverse model. Continuous – real output; dashed – TS fuzzy model

Acoustic TS fuzzy model Inverse TS fuzzy model

RMS (dB) 3.75 2.46
VAF 87.3% 92.1%
Table 2: Acoustic and inverse model accuracy measures

11
 −10 −10
−20 −20
Power Spectrum Magnitudde [dB]

Power Spectrum Magnitudde [dB]

−30 −30
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
−100 −100
200 400 600 800 1000 200 400 600 800 1000
Frequency [Hz] Frequency [Hz]

(a) Classic controller. Continuous – real output; dashed (b) Inverse fuzzy control. Continuous – real output;
– controller. dot-dashed – controller.

Figure 9: Tonal excitation control.

6.3 Control
This section presents the experimental control results. The excitation is a tonal signal of 370 Hz
with harmonics. The classic control loop shown in Fig. 1 and the control loop proposed in this
paper and shown in Fig. 4 are both applied to cancel the noise. Figure 9 depicts the obtained
results for both controllers. Both controllers attenuate the dominant frequency (370 Hz) for
which they were designed. However, some harmonics are amplified by the classic controller,
see Fig. 9a. On the other hand, the inverse fuzzy controller do not amplify any harmonics as
shown in Fig. 9b. The global sound reduction is obtained by the inverse fuzzy model-based
controller.

7 CONCLUSIONS
This paper presented an active noise controller based on fuzzy acoustic modeling of the envi-
ronment where one wants to cancel noise and proposed a direct derivation of an inverse model
controller to be applied in a feedforward strategy. The control performance are compared to a
classical filter controller. The proposed fuzzy controller is able to accomplish a global noise
reduction.
To control enclosed sound fields it is necessary to use several devices: multiple micro-
phones, since the attenuation at one point usually implies the amplification at another points;
and multiple loudspeakers, to control all possible modes in an efficient way. In classic linear ap-
proaches, the increase in the number of variables represent an exponential increase in number
of filters and controllers, and consequently an increase in the control loops complexity. Us-
ing fuzzy models, this complexity can be reduced, since a single model can represent multiple

12
inputs and outputs, and the control loop will be exactly the same.
Future research will focus the extrapolation of this type of controller to multiple-input-
multiple-output systems, and its consequences in terms of computational burden.

Acknowledgements
This work was supported by PRAXIS XXI, ref. PRAXIS/C/EME/12124/1998, Fundação para
a Ciência e Tecnologia, Ministério da Ciência e da Tecnologia, Portugal.

References
[1] R. Babuška. Fuzzy Modeling for Control. Kluwer Academic Publishers, Boston, 1998.

[2] R. Babuška, J.M. Sousa, and H.B. Verbruggen. Inverse fuzzy model based predictive
control. In D. Driankov and R. Palm, editors, Advances in Fuzzy Control, pages 129–154.
Physica-Verlag, Heidelberg, New York, 1998.

[3] D.E.Gustfason and W.C.Kessel. Fuzzy clustering with a fuzzy covariance matrix. In In
Proceeding IEEE CDC, pages 761–766, San Diego,CA, USA, 1999.

[4] S.J. Elliott. Signal processing for active control. Academic Press, London, UK, 2000.

[5] Stephen J. Elliott. Active control of nonlinear systems. In Active 99, USA, 1999.

[6] Lawrence E. Kinsler, Austin R. Frey, Alan B. Coppens, and James V. Sanders. Funda-
mentals of Acoustics. John Wiley and Sons, New York, USA, fourth edition, 2000.

[7] M.Fischer and R.Isermann. Robust hybrid control based on inverse fuzzy process models.
In In Proceeding IEEE International conference on fuzzy systems, pages 1210–1216, New
Orleans, USA, 1996.

[8] P.A. Nelson and S.J. Elliott. Active Control of Sound. Academic Press, UK, 1999.

[9] M. Setnes, R. Babuška, U. Kaymak, and H. R. van Nauta Lemke. Similarity measures
in fuzzy rule base simplification. IEEE Transactions on Systems, Man and Cybernetics –
Part B: Cybernetics, 28(3):376–386, 1998.

[10] C.A. Silva, J.M. Sousa, M.A. Botto, and J.M.G. Sá da Costa. Active acoustic noise iden-
tification using fuzzy modeling techniques. In ISCSV7 - Seventh International Congress
on Sound and Vibration, pages 3213–3220, Garmisch-Partenkichen, Germany, july 2000.

[11] C.A. Silva, J.M. Sousa, and J.M.G. Sá da Costa. Model-based active noise control using
fuzzy models. In Controlo2000 - Fourth Portuguese conference on control, pages 377–
382, Guimarães, Portugal, october 2000.

13
[12] M.O. Tokhi and R.R. Leitch. Active Noise Control. Oxford Science Publications, engi-
neering science series - 29 edition, 1992.

[13] T.Takagi and M.Sugeno. Fuzzy identification of systems and its application to modeling
and control. IEEE Transactions on Systems, Man and Cybernetics, 15(1):116–132, 1985.

[14] X. Zeng and M. Singh. Approximation theory of fuzzy systems - mimo case. IEEE Trans.
Fuzzy Systems, 3(2):219–235, 1995.

14