Enhanced inverse model LMS algorithm for active control of

harmonic response clusters

Mingfeng Li
a)

Jie Duan
b)
Teik C. Lim
c)
School of Dynamic Systems
College of Engineering and Applied Science
University of Cincinnati
Cincinnati, OH 45220

Harmonic responses can be seen widely in many applications, such as vehicle powertrain
noise, gear vibration, and other rotating machineries. Hence, to develop an effective system
for harmonic control is highly desirable. For single frequency control, it can be easily
suppressed by most active control systems. However, when a cluster of harmonic responses
that spreads over a wide frequency range or contains rapidly varying frequencies, the
traditional filtered-x least mean square (FXLMS) algorithm may not be able to achieve
significant enough reduction over a wide frequency range or track variations in frequency
due to the large dynamic range of secondary path in that wide spectral range. In this study,
an enhanced algorithm, namely the inverse model least mean square algorithm (IMLMS),
which utilizes the inverse model of the secondary path to minimize the effect of the
secondary path dynamics on the algorithm convergence, is proposed to improve the
convergence property and tracking ability of the active control system. Numerical
simulation using synthesized signals was performed. Results show significant improvement
on the convergence of the system and in turn achieving more reductions in the responses
over a wide frequency range by the proposed scheme.

a)
email: [email protected]
b)
email: [email protected]
c)
email: [email protected]

1 INTRODUCTION

Active noise control (ANC) is to abate the noise by introducing anti-phase sound waveform
to counteract the original sound response at target locations
1
. In the past decades, active noise
control (ANC) is receiving increasing number of attention in both academia and industries. The
successful applications of ANC technology are also increasing in the recent years. Based on the
problems that ANC system is used, two categories, namely broadband noise control and
harmonic control, can be classified. Harmonic response is prevalent in the rotating machineries,
such as motors, fans, turbines, gearboxes, compressors, pumps and powertrains, and as well as
systems consisting of electronic and electromechanical components, such as MRI scanners, hard
disk drives, transformers, and computers. They often produce acoustic noise that can be
annoying and sometimes loud enough to be harmful to human hearing. Hence, controlling
harmonic responses is very desirable for many applications. For some applications, the frequency
of the harmonics may vary with the time. Hence, one of the most important factors of the
designed active control system might be the convergence property or tracking ability of the
algorithm.
The traditional active control system for harmonic control is essentially a feedforward type
where the reference signal is created according to the frequency information of the disturbance
such as the speed of shaft in rotatory system, as shown in Fig. 1. As early as in 1977, Glover
used the sum of the sinusoidal components as the reference signal
2
. In 1985, Elliot and
Darlington synchronously sampled the interference to improve the efficiency of the harmonic
controller
3
. Later, Bodson and his group proposed an indirect approach to control the harmonic
responses, where the frequency is unknown
4,5
.

Fig. 1 - Traditional feedforward control system for suppressing harmonic disturbance

Convergent speed of the system will affect the system performance, especially for
frequency-varying cases such as engine ramp-up. One main factor affecting the convergent speed
is the varying power of filtered reference signal. The convergence property of the algorithm is
controlled by step size of the FXLMS algorithm. However, the step size has an upper limit which
is determined by power of filtered reference signal and the secondary path dynamics
1
. The
secondary path is defined as the transfer function from the input of the control source to the
output of the error sensor. In most of ANC applications, the control source is typically a speaker
while the error sensor is a microphone. Hence, the dynamics of secondary path is normally
highly dynamic system. In 2008, the eigenvalue equalization method, proposed by Sommerfeldt
et al.
6,7
, provided a certain degree of improvement for suppressing harmonic response. However,
it did not address the converging speed issue due to the phase delay of secondary path.
To overcome the convergence problem existing in most of control systems, a new active
control system that is effective in treating harmonic signals is proposed. The proposed algorithm
does not employs the convolution calculation to obtain the filtered reference signal, and uses the
least mean square (LMS) formulation instead of the more complex FXLMS approach to increase
the overall convergent speed over a broad frequency range.

W(z)
u(n)
y(n)
r(n)
Sine Wave
Generator
FIR
Rotation Speed
or disturbance
frequency
+
+
LMS
d(n) Disturbance
Noise
e(n)
2 INVERSED MODEL LMS ALGRITHM

Since the basic LMS does not need a filtered reference signal to update the adaptive filter
weights, there is no need for the convolution operation. Also, the converging speed of the basic
LMS algorithm only depends on the power of reference signal. However, basic LMS algorithm
can only be applied to cases where no secondary path exists. In most ANC applications, there is
always a secondary path in the system. Hence, to apply the basic LMS algorithm in an active
control system, we have to find a way to eliminate the effect of the secondary path. One
approach is to add an inverse model of the secondary path after the controller as shown as Fig. 2.
However, for most cases, the secondary path is a non-minimum phase system. Hence, a stable
and casual inverse model does not exist for broadband noise control. Fortunately, for harmonic
control, this is not an issue because a non-casual filter can be used, which allows the system
shown in Fig. 2 to be implemented for harmonic response control.

Fig. 2 - Applying the basic LMS algorithm by using an inverse model of the secondary path after
the controller.

Two different implementations are developed as shown in Figs. 3 and 4. The first
implementation is shown in Fig. 3. The output of the controller, u(n), is phase shifted and
multiplied by a gain. The values of the phase shift and the gain can be calculated from the
measured response of the secondary path. The response of the inverse model of the secondary
path can be expressed as


0
1/ ( ) 1/( ) H j A jB = + (1)

where
0
is the frequency of the reference signal, A and B are real and imaginary parts of the
secondary path response, which can be obtained through a system identification. Hence the phase
shift and gain g in Fig. 3 can be expressed as arctan( / ) B A = and
2 2
1/ g A B = + .
The second implementation is shown in Fig. 4. We know that the output of control filter
W(z) is still a sinewave signal but with phase shifted and amplified or attenuated of the reference
signal. Assuming the reference signal
0
( ) cos( ) r n n = , then the output of control filter will be
0 0 0
( ) cos( ) u n g n = + , when g
1
and
1
are the gain and phase of the control filter. The signal
u(n) is needed to filtered by the inverse model of secondary path. That means the output the
W(z)
u(n)
y(n)
r(n)
Sine Wave
Generator
FIR
Rotation Speed
or disturbance
frequency
+
+
LMS
d(n) Disturbance
Noise
e(n)
1/H(z)
inverse model should be
0 0 0
cos( ) g g n + + . If we know
0 0 0
sin( ) g n + , we can easily
obtain the output after inverse model, which is
0 0 0 0 0 0 0 0 0
cos( ) cos( ) sin( ) g g n Ag n Bg n + + = + + + (2)
Fortunately,
0 0 0
sin( ) g n + can be easily obtained through filtering
0
sin( ) n by control filter
W(z) that is shown in Fig. 4.


Fig. 3 - First implementation: using phase shifting and multiplying a gain to the output of the
controller, u(n), to eliminate the secondary path effect at that frequency.


Fig.4 - Second implementation: tuning the gains of the cosine and sine components to shift the
phase and adjust the gain.

H(z)
u(n)
y(n)
r(n)
Sine Wave
Generator
W(z)
+
+
LMS
d(n) Disturbance
Noise
e(n)
A
Rotation Speed
or disturbance
frequency
B W(z) 90
copy
H(z)
u(n)
y(n)
r(n)
Sine Wave
Generator
W(z)
+
+
LMS
d(n) Disturbance
Noise
e(n)
Phase
shift
Gain
Secondary Path
response
Frequency
Rotation Speed
or disturbance
frequency
3 NUMERICAL SIMULATION

To demonstrate the performance of the proposed system, the algorithms are applied to
control a set of steady-state and time-varying harmonics in numerical simulation. The used
secondary path in the simulation is shown in Fig. 5, which is a measured dynamic response.
Simulation results for these two implementations compared to the traditional FXLMS approach
are shown in Figs. 6-10.
Figure 6 is the time history of controlled response for multiple steady-state harmonics at
constant frequencies from 100 Hz to 900 Hz with 100 Hz increment using the first ANC
implementation (FIR adaptive filter) proposed. The top plot of Fig. 6 is the controlled response
of the proposed control system. The bottom plot of Fig. 6 is the controlled response using the
traditional FXLMS algorithm. Obviously, the proposed system can yield faster convergence
speed than the traditional FXLMS algorithm. Figure 7 shows the frequency spectrum and the
corresponding reduction of the controlled response using the last one second of data from Fig. 6.
Results show clearly that over 50 dB reductions are obtained for all harmonic components when
it is applying the proposed algorithm for the harmonic control. For the traditional system, much
less reductions are achieved at some high frequency components, especially for 800 Hz and 900
Hz. Even at the last second of the simulation, less than 20 dB reduction at 800 Hz can be
obtained. For the 900 Hz, only few dB reductions is seen. Hence, the overall converging speed of
the traditional system is much slower than the new system due to the slow convergence at high
frequencies. The reason is the low responses of secondary path at high frequencies as shown in
Fig. 5.
Figures 8-10 are the simulation results using the second ANC implementation proposed. Fig.
8 is the time history of controlled response for multiple steady-state harmonics at constant
frequencies of from 100 Hz to 900 Hz with 100 Hz increment. The top plot of Fig. 8 is the
controlled response of our proposed system. The bottom plot of Fig. 8 is the controlled response
using the traditional FXLMS algorithm. Figure 9 shows the response spectrum and
corresponding reduction of the controlled response using the last one second of data from Fig. 8.
From the results, similar conclusion as the first implementation can be drawn. In the Fig. 10, it
shows the simulation results of the swept sine case with the frequency linearly changing from
200 Hz to 800 Hz in 5 seconds. The black line represents the control results for the traditional
FXLMS algorithm. Red line is the results from the proposed system. Similarly, the results show
significant improvement achieved by the proposed system as compared to the traditional one.


4 CONCLUSIONS

In this paper, an enhanced algorithm, namely the inverse model least mean square algorithm
(IMLMS) was proposed to improve the control performance for the harmonic responses by using
the inverse model of the secondary path to minimize the effect of the secondary path dynamics
on the algorithm convergence. Numerical simulation was performed and compared between
traditional FXLMS algorithm and proposed systems. Results show significant improvement on
the convergence of the system and in turn achieving more reductions in the responses over a
wide frequency range by the proposed scheme.


5 REFERENCES

1. S. M. Kuo and D.R. Morgan, Active Noise Control Systems: Algorithm and DSP
Implementation, J ohn Wiley & Sons, Inc, New York, (1996).

2. J . R. Glover J R., Adaptive noise canceling applied to sinusoidal interferences, IEEE
Transaction on Acoustics, Speech, and Signal Processing, 25(6), 484-491, (1977).

3. S.J . Elliott and P. Darlington, Adaptive cancellation of periodic, synchronously sampled
interference, IEEE Transaction on Acoustics, Speech, and Signal Processing, 33(3), 715-
717, (1985).

4. M. Bodson and S.C. Douglas, Adaptive algorithms for rejection of sinusoidal interferences
with unknown frequency, Automatic, 33(12), 2213-2221, (1997).

5. B. Wu and M. Bodson, Multi-channel active noise control for periodic sources-indirect
approach, Automatic, 40, 203-212, (2004).

6. S.D. Sommerfeldt, J . Blotter, B.M. Faber, Secondary path modeling for active noise control,
US Patent, 2008/0144853 A1, (2008).

7. J .K. Thomas, S. Lovstedt, J .D. Blotter, S.D. Sommerfeldt, and B. Faber, Eigenvalue
equalization filtered-x (EE-FXLMS) algorithm applied to the active minimization of tractor
noise in a mock cabin, Noise Control Engineering Journal, 56(1), 2534, (2008).


0 100 200 300 400 500 600 700 800 900 1000
-50
-40
-30
-20
-10
0
10
M
a
g
n
i
t
u
d
e

(
d
B
)
0 100 200 300 400 500 600 700 800 900 1000
-200
0
200
Frequency (Hz)
P
h
a
s
e

(
D
e
g
)

Fig.5 - Frequency response of secondary path used in simulation
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
-5
0
5
P
r
o
p
o
s
e
d

L
M
S
_
I
n
v
e
r
s
e

S
y
s
t
e
m
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
-5
0
5
Time (Sec)
T
r
a
d
i
t
i
o
n
a
l

F
X
L
M
S

S
y
s
t
e
m
Fast convergence
Slow convergence

Fig. 6 - Time history of controlled response for multiple harmonics at constant frequencies of
100~900 Hz. (Top plot, proposed system; bottom plot, traditional FXLMS algorithm)


0 100 200 300 400 500 600 700 800 900 1000
-60
-40
-20
0
M
a
g
n
i
t
u
d
e

(
d
B
)
100 200 300 400 500 600 700 800 900
0
20
40
60
80
Frequency (Hz)
R
e
d
u
c
t
i
o
n

(
d
B
)
More reduction
Less reduction

Fig. 7 - Frequency spectrum of controlled response for multiple harmonics at frequencies of
100~900 Hz using the last one second of simulation data. (Keys: blue line, baseline
response; red line, proposed new system; black line, traditional FXLMS algorithm)


0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
-5
0
5
P
r
o
p
o
s
e
d

L
M
S
_
I
n
v
e
r
s
e
d

S
y
s
t
e
m
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
-5
0
5
Time (Sec)
T
r
a
d
i
t
i
o
n
a
l

F
X
L
M
S

S
y
s
t
e
m
Slow convergence
Fast convergence

Fig. 8 - Time history of controlled response for multiple harmonics at constant frequencies of
100~900 Hz. (Top plot, proposed system; bottom plot, traditional FXLMS algorithm)

0 100 200 300 400 500 600 700 800 900 1000
-60
-40
-20
0
M
a
g
n
i
t
u
d
e

(
d
B
)
100 200 300 400 500 600 700 800 900
0
20
40
60
80
Frequency (Hz)
R
e
d
u
c
t
i
o
n

(
d
B
)
Less reduction More reduction

Fig. 9 - Frequency spectrum of controlled response for multiple harmonics at constant
frequencies of 100~900 Hz using the last one second of simulated data. (Keys: blue
line, baseline response; red line, proposed new system; black line, traditional FXLMS
algorithm)
200 300 400 500 600 700 800
-25
-20
-15
-10
-5
0
Frequency (Hz)
M
a
g
n
i
t
u
d
e

(
d
B
)

Fig. 10 - Frequency spectrum of controlled response for swept sine case with frequency
changing from 200 Hz to 800 Hz in 5 seconds. (Keys: blue line, baseline response;
red line, proposed new system; black line, traditional FXLMS algorithm)