Proceedings of the 18th World Congress

The International Federation of Automatic Control

Milano (Italy) August 28 - September 2, 2011

Disturbance Rejection through LPV Gain-Scheduling Control

with Application to Active Noise Cancellation
Pablo Ballesteros, Christian Bohn
Institute of Electrical Information Technology, Clausthal University of Technology, Germany,
(e-mail: ballesteros@iei.tu-clausthal.de, bohn@iei.tu-clausthal.de).

Abstract: A design method for the design of discrete-time gain-scheduling controllers for the rejection of
disturbances with time-varying dynamics is presented. The disturbance is modeled as the output of a
linear parameter-varying system in linear-fractional-transformation form. The work is motivated by the
rejection of harmonic disturbances with time-varying frequencies, a problem that arises in active noise
and vibration control. The design method is described in detail and experimental real-time results
obtained with an active noise control headset are presented. Over existing approaches (such as adaptive
filtering or gain-scheduled observer-based state feedback) the proposed method has the advantage that it
leads to a stable closed-loop system even for arbitrarily fast changes of the disturbance frequencies.

1. INTRODUCTION or by interpolating or switching between a set of gains that

A common control problem treated in active noise and
vibration control (ANC/AVC) is the rejection of harmonic
disturbances with time-varying, but known (measured)
frequencies. Such disturbances occur, for example, in
applications where rotating machinery operate with varying
speeds, for example, in automotive vehicles.
For the rejection of harmonic disturbances different
approaches are possible. Often, adaptive feedforward
controllers based on the Filtered-x LMS (FxLMS) algorithm
(see, e.g., Kuo and Morgan 1996) are used. The FxLMS
algorithm works well in practice but convergence speed and
tracking performance might pose some problems. Since the
controller results from the adaptation, an off-line analysis of
the closed-loop behavior is difficult. Also, to date, only
approximate stability results for the FxLMS algorithm seem
available (Feintuch et al. 1993, Kuo and Morgan 1996).
Another solution to reject disturbances is feedback control. It
is well known, that for good disturbance rejection, the
feedback controller has to include a model of the disturbance
[this is the internal model principle; Francis and Wonham
(1976)]. Basically, this leads to a controller that, if analysed
in the stationary case, contains the poles of the disturbance
model that then show up as zeros in the closed-loop transfer
function. For the rejection of harmonic disturbances, this
means that the controller has infinite gain at the frequencies
that should be suppressed. One way to achieve this is to use
an observer that estimates the plant states and also the states
of a disturbance model. The estimated states of the
disturbance can then be used to cancel the disturbance. For
harmonic disturbance rejection, this method has been used,
for example, by Bohn et al. (2003, 2004) and Kinney and
Callafon (2007). The observer used is basically a time-
varying and possibly scheduled version of the spectral
observer (Hostetter 1980). The time-varying observer gain
can be calculated on-line (for example, as the Kalman gain)
are precomputed for fixed frequencies (Bohn et al. 2003,
2004; Kinney and Callafon 2007). In the first case, stability
can be guaranteed but the computational effort is increased
(due to the on-line calculation of the covariance matrix). In
the latter case, the computational effort is lower (only table
lookup and interpolation operations) but stability is not
guaranteed. Another alternative is to use observer-based
feedback control but to schedule the controller instead of the
observer (Kinney and Callafon 2006a, 2006b and Heins et al.
2011). The continuous-time interpolation approach presented
by Kinney and Callafon (2006a) guarantees stability.
In a norm-optimal control framework, good disturbance
rejection can be achieved by using a disturbance model as a
weighting function. This approach is pursued in this paper.
The problem of rejecting harmonic disturbances with time-
varying frequencies can be seen as a special case of a more
general scenario, namely the rejection of disturbances with
linear dynamics that depend on time-varying parameters. In
this paper, disturbances are considered that can be adequately
described as the output yd of a linear parameter-varying
state-space model
xd, k 1  A d ( pk ) xd, k  Bd, w wd, k , (1)
yd, k  C d, y xd, k , (2)
where the parameter vector pk determines the time-varying
nature of the disturbance dynamics. It is assumed that this
model can be converted to linear fractional transformation
form (see Fig. 1)
xd, k 1  A d, 0 xd, k  Bd,  w , k  Bd, w wd, k , (3)
q , k  C d,  xd, k , (4)
yd, k  C d, y xd, k , (5)
w , k  θk q , k , (6)

978-3-902661-93-7/11/$20.00 © 2011 IFAC 7897 10.3182/20110828-6-IT-1002.00830

18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

θ θ
w q w q
wd Gd  z  yd w q
Gz
u y
Fig. 1. Disturbance model in LPV-LFT form.
where the time-varying disturbance characteristics are now K z
captured by the time-varying parameter θ .
q w 
If this disturbance model is combined with the plant model,
the gain-scheduling design method for LPV-LFT systems can θ
be used (Apkarian and Gahinet 1995). This results in a
controller that uses the time-varying parameter θ as a Fig. 2. LPV gain-scheduling system.
scheduling variable and consequently in the overall control
system shown in Fig. 2. where yd is the output of the disturbance model given by
eqs. (3) – (6). For simplicity, only single-output systems with
A similar approach was used by Du and Shi (2002) and Du et one single disturbance and a single control input are
al. (2003) for the rejection of a single harmonic disturbance
considered. The extension to general multivariable systems is
with time-varying frequency. The disturbance was modeled
straightforward.
as the output of an LPV model in polytopic form and tested
in a simulation example. Kinney and Callafon (2006a) used a For norm-optimal control design, performance inputs and
similar disturbance description (polytopic LPV) but included outputs are defined via additional weighting functions. Since
the disturbance model directly in the controller and observed the focus is on disturbance rejection, weighting functions
only the plant states. Both Du and Shi (2002) and Du et al
 AW y B Wy 
(2003) and Kinney and Callafon (2006a) designed W yp    (9)
continuous-time controllers (that were only tested in  C Wy D Wy 
 
simulation). Working in continuous time, however, poses
problems in the real-time implementation since the controller and
has to be discretised in each sampling instant. This results in
 AWu B Wu 
a high computational load (calculation of a matrix Wu p    (10)
exponential). An approximate discretisation leads to  CWu D Wu 
 
frequency distortion that cannot be tolerated if harmonic
disturbances have to be cancelled. Therefore, in this paper, a are used for the output and the control signal, respectively.
discrete-time design procedure is used. The generalized plant including the plant, the disturbance
model and the weighting functions is shown in Fig. 3. The
The design method presented here leads to a fairly simple state space representation of this generalized plant G(z) is
controller structure (see Fig. 2). The resulting gain- given by
scheduling controller has a fairly low computational load
 Ap Bd C d 0 0 0 0 Bu 
(compared, for example, to the covariance update required in  
a time-varying state estimator) and implementation is  0 A d, 0 0 0 Bd,  Bd 0 
 
straightforward. Also, closed-loop stability is guaranteed for  x k 1   B Wy C p 0 A Wy 0 0 0 0
  xk 
the whole range of parameter variations specified in the    0 0 0 A Wu 0 0 B Wu   w , k 
 q , k      . (11)
design, even if the parameters vary arbitrarily fast.  qk   0 C d,  0 0 0 0 0   wd, k 
   
The remainder of this paper is organized as follows. In  yk   D W C p 0 CWy 0 0 0 0  u k 
   y  
Sec. 2, it is described how the disturbance model and the  0 0 0 CWu 0 0 D Wu 
 
plant model are combined to form the generalized plant. The  Cp 0 0 0 0 0 0 
 
design procedure is briefly outlined in Sec. 3. In Sec. 4, the
modeling of harmonic disturbances is discussed and in Sec. 5 3. CONTROL DESIGN
experimental real-time results are presented. For the real-time
results, an active noise control headset is used and two Although the design method is described in detail in
harmonic disturbances are considered. In Sec. 6, some Apkarian and Gahinet (1995), it is briefly reviewed here. The
conclusions are given. starting point for the control design is the description of the
generalized plant
2. SYSTEM MODELING
 x k 1   A B Bw Bu   xk 
The plant is represented by the state space model     
 q , k    C D D w D u   w , k 
. (12)
xp, k +1  A p xp, k  B u yd, k  Bu u p, k , (7)  qk   C q Dq Dqw Dqu   wd, k 
    
yp, k  C p xp, k , (8)  yk   C y D y Dyw D yu   u k 

7898
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

w θ q   X 1 A0 B0 0 
 T 
X C 0T 
Gd  z   B
A
ψ   0T
0
0
 L0 D0T 

, P  B T 
0 0 D 12T , (22)
w wd  0

 0 C D0  J 0 
yd  0

u up yp
 
+
Gp  z  W yp  z   Q  0 C D 21 0 , (23)

q
  A 0  0 B Bw    0 Bu 0
 A0    , B0   , B   , (24)
Wup  z    0 0  0 0 B I 0 0
G z y  0 0  0 I 0 0 0 
     
C 0   C 0  , C   C y 0  , D0   0 D D w  , (25)
Fig. 3. Plant and disturbance in LPV-LFT form. C 0   0 0  0 D Dqw 
 q   q

To compute the controller, first, matrices N R and N S are

0 0 I 0 0 0 
obtained as      
D12   0 D u 0  , D 21   0 Dy Dyw  , (26)

N R  null BuT DTu T
Dqu 
0 , (13) 0 D
 qu 0  I
 0 0 

NS  null  C y Dy Dyw 0 , (14) and

and an upper bound  on the H  -norm of the closed-loop L L2   L 0 1

L   T1  , L0    , J0  L (27)
system is specified. Then, the system of LMIs given by  L2 L3   0 I 
 ARAT  R ARCT ARCqT B Bw  is solved for Ω and the state-space matrices of the controller
  are extracted from Ω , which is
 C RA
T
 J 3  C RCT C RC T
q D D w 
T
N R C q RA T
C q RCT  I  C q RCqT Dq Dqw  N R  0, (15)
   AK BK 
 BT
DT DqT  L3 0  Ω . (28)
  CK DK 
 BwT
DTw T
Dqw 0  I   

 AT SA  S AT SB AT SBw CT C qT  4. APPLICATION TO THE REJECTION OF

  HARMONIC DISTURBANCES
T
 B SA  L3  BT SB BT SBw DT DqT 
T T 
NS

BwT SA BwT SB  I  BwT SBw DTw Dqw N S  0,

(16) The method described in the previous section is now applied
 C D D w  J 3 0  to compute a controller that rejects harmonic disturbances
 
 Cq Dq Dqw 0  I  with time-varying frequencies.
and A single harmonic disturbance with frequency f (i ) can be
R I  L3 I  modeled as the output of the state space model
   0,    0, (17)
I S  I J3  xd,( i )k 1  A(di ) xd,( i )k  Bd,( i )w wd,( i )k (29)
is solved for the symmetric matrices R and S and the diagonal yd,( i )k 1  C d,( i )y xd,( i )k 1 (30)
matrices J 3 and L3 . From J 3 and L3 , matrices L1 and L2
with
are determined that satisfy
0 1  (i ) 1 (i )
L3  J 31  LT2 L11 L2 (18) A(di )   (i ) 
, Bd, w    , C d, y  1 0 , (31)
 1 a  1
and from R and S, matrices M and N of minimal
dimensions are computed that satisfy and

MN T  I  RS . (19) a (i )  2 cos ( 2 f (i ) T ) , (32)

From R, S, M and N , the matrix X is obtained as where T is the sampling time It should be noted that this
1
model is only correct for a constant frequency. For a time-
 S I  I R  varying frequency, a correct state transition matrix would be
X  T   . (20)
N 0  0 MT 
 cos (2 f k(i )T ) sin (2 f k(i )T ) 
A(d,i )k   , (33)
Finally, the LMI
  sin (2 f k T ) cos (2 f k T ) 
(i ) (i )

ψ  QT ΩT P  P T Ω Q  0 (21) 2 f k(i )   k(i)1   k(i ) (34)

with
where  (i )
k is the phase angle of the disturbance.

7899
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

Nevertheless, the “incorrect” model is used here since it is 5. EXPERIMENTAL RESULTS

slightly simpler. This does not affect the performance of the
The controller obtained from the design procedure of Sec. 3
system for constant disturbance frequencies or the closed-
with the disturbance model of Sec. 4 is validated
loop stability. It might affect the disturbance attenuation for
experimentally on an active noise control headset (Sennheiser
fast changes in the disturbance frequency. In most practical
PXC 300). The headset has two microphones placed on the
applications with time-varying frequencies, however, the
ear cups of the headset. The setup of the systems is
measured frequency will not correspond exactly to the
schematically shown in Fig. 4. The objective is to cancel the
instantaneous frequency due to measurement delays and the
harmonic disturbances produced by an external loudspeaker
transmission of the disturbance to the plant. This means that
with the loudspeakers of the headset.
the measured frequency will never satisfy eq. (34). It is then
unclear whether using the correct model would result in The control algorithm is implemented on a rapid control
better performance. prototyping unit (dSpace MicroAutoBox). An anti-aliasing
(i ) (i )
filter is used for the output signal and a reconstruction filter
If the frequency varies between f min and f max , the parameter for the control input. Two independent controllers are
a(i ) will vary between a(min
i)
and a (max
i)
and can be written as implemented for both sides of the headset.
a (i )  a0(i )  a1(i ) (i ) , (35) The transfer function between the output signal and the input
signal of the control unit is the plant Gp(z) (usually called
where a0(i ) and a1(i ) are real constants and  (i )  [1, 1] . The secondary path in active noise and vibration control). To
model for the harmonic disturbance can then be written in compute the controller, this transfer function is required. For
LPV-LFT form as obtaining this transfer function, the system is excited with a
multisine test signal, output and input signals are recorded.
xd,( i )k 1  A(d,i )0 xd,( i )k  Bd,( i ) w( i,)k  Bd,( i )w wd,( i )k , (36) The transfer function can be estimated using a standard
q(i,)k  C d,( i ) xd,( i )k , (37) black-box technique (oe). All usual methods (arx, oe, n4sid)
resulted in models that were suitable for control design. This
yd,( i )k  C d,( i )y xd,( i )k , (38) was done for both sides of the headset. Since the transfer
functions were almost identical, the same transfer function
w( i,)k   k( i ) q( i,)k , (39) was used for the control design and consequently, the same
with controllers were used for both sides. The experimental results
presented are for the left side. The transfer function is of 12th
0 1  (i )  0  (i ) order and the controller of 17th order.
A(d,i )0   (i ) 
, Bd,    ( i )  , C d,    0 1 . (40)
  1 a0  a1  The disturbance is the sum of two independent harmonic
More than one frequency can be considered by adding the disturbances in the frequency range from 90 Hz to 110 Hz
outputs of individual single-frequency models. The overall and 110 Hz to 140 Hz, respectively. A first-order low pass
disturbance model for N frequencies is then given by filter is used as weighting function for the output. The control
signal is weighted by a constant gain. A sampling frequency
xd, k 1  A d, 0 xd, k  Bd,  w , k  Bd, w wd, k , (41) of 1 kHz was chosen such that the Nyquist frequency of 500
Hz is well above the highest disturbance frequency.
q , k  Cd,  xd, k , (42)
Design results and experimental data are presented in Figs. 5-
yd, k  Cd, y xd, k , (43) 7. Fig. 5 shows results for constant disturbance frequencies of
w , k  θk q , k , (44) 90 Hz and 140 Hz and Fig. 6 results for constant disturbance
frequencies of 100 Hz and 120 Hz. From the amplitude
with frequency responses excellent disturbance rejection is
 A(1) 0   Bd,(1) 0  expected, which is confirmed by the measured microphone
d, 0
    signals.
A d, 0    , Bd,     , (45)
 0 A(d,N0)   0 (N ) 
Bd,  
 
 Bd,(1)w 0    (1) 0 
   
Bd, w    , θ    , (46)
 0 (N ) 
Bd, w   (N ) 
 
  0
 Cd,(1) 0 
 
Cd,     , Cd, y  Cd, y  Cd, y .
(1) (N )
  (47)
 0 Cd,( N ) 

The augmented plant G(z) can be obtained as described in
Sec. 2 and the controller K ( z ) using the procedure of Sec. 3. Fig. 4. Block diagram of the ANC System.

7900
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011

The frequency responses in Figs. 5 and 6 show that the 4

disturbance attenuation at the disturbance frequencies leads to Closed Loop
some disturbance amplification for frequencies below and Open Loop
3

Amplitude [Pa / V]
above the disturbance frequency. This is due to the
“waterbed” effect (Bode’s sensitivity integral), sometimes
2
also called “spillover” (Hong and Bernstein 1998), although
this must be distinguished from “modal spillover” (Hansen
2001). Whether this is tolerable in a practical application 1
depends on the spectral content of the background noise. This
might be a disadvantage of the feedback approach over 0
0 100 200 300 400 500
feedforward approaches, which do not necessarily lead to Frequency [Hz]
disturbance amplification.
In Fig. 7 the behavior for a disturbance with time-varying 0.05
frequencies is shown. The disturbance is the sum of two sine
sweeps with frequencies linearly increasing from 90 Hz to
110 Hz and from 110 Hz to 140 Hz, respectively, over 10

Pressure [Pa]
seconds. The comparison of the pressure measured at the 0
microphone for the open-loop and closed-loop case show that
excellent disturbance attenuation is achieved. The control
system also performed well for sweeps with duration of 5
seconds. For even faster sweeps, the system remained stable
but did not achieve satisfactory disturbance attenuation. This -0.05
0 2 4 6 8
could be a consequence of using an “incorrect” disturbance Time [s]
model, as pointed out in Sec. 4. This will be investigated in
Fig. 6. Results for fixed disturbance frequencies of 100 Hz
future work.
and 120 Hz: Open-loop and closed-loop amplitude frequency
responses (top) and pressure measured at the microphone
(bottom), the controller is switched on at approx. t  1 sec
5 and off at approx. t  7 sec .
Closed Loop
4 Open Loop
140
Amplitude [Pa / V]

ff(1)
1
3 130 ff(2)
2
Frequency [Hz]

2 120

1 110

0 100
0 100 200 300 400 500
Frequency [Hz]
90
0 2 4 6 8 10
Time [s]

0.05
0.15
Pressure [Pa]

0.1
0
0.05
Pressure [Pa]

0
-0.05
-0.05

0 2 4 6 8 -0.1
Time [s]
-0.15
0 2 4 6 8 10
Fig. 5. Results for fixed disturbance frequencies of 90 Hz and Time [s]
140 Hz: Open-loop and closed-loop amplitude frequency
responses (top) and pressure measured at the microphone Fig. 7. Results for a disturbance with time-varying
(bottom), the controller is switched on at approx. t  1 sec frequencies: Variation of the frequencies (top) and pressure
and off at approx. t  7 sec . measured at the microphone (bottom) in open loop (light
gray) and closed loop (black).

7901
18th IFAC World Congress (IFAC'11)
Milano (Italy) August 28 - September 2, 2011
6. CONCLUSIONS
Motivated by the rejection of harmonic disturbances with
time-varying frequencies in active noise and vibration
control, a general design procedure to reject disturbances that
can be described in LPV-LFT form is presented. The
disturbance model, the plant and performance weighting
functions are combined to a generalized plant and the gain-
scheduling design method of Apkarian and Gahinet (1995) is
used. The control design method results in a controller that
uses the disturbance frequency as a scheduling variable.
The approach exhibits similarities to the continuous-time
polytopic LPV approach presented by Du and Shi (2002) and
Du et al. (2003). The main differences are that the design is
carried out in discrete time, which avoids problems in the
implementation of the control algorithm (discretization at
each sampling instant and frequency distortion resulting from
approximate discretization) and that an LPV-LFT form is
used. Also, here a disturbance consisting of two harmonics is
considered and real results obtained for an ANC headset are
presented. Results for AVC and ANC are presented in
Ballesteros and Bohn (2011).
The experimental results show that excellent disturbance
rejection is achieved for constant frequencies and for time-
varying frequencies, even when the frequencies change fairly
rapidly. An advantage over adaptive filtering and heuristic
controller interpolation approaches is that closed-loop
stability is guaranteed. Since the controller can reject
disturbances in a specified frequency range, the design
procedure might be relevant even for cases where the
frequencies are constant. For example, the algorithm could be
used to built a programmable active noise control headset
where the user could manually set and reset the fixed
frequencies that should be cancelled.
Future work will focus on the rejection of large numbers of
harmonics [seventeen harmonics were cancelled in Bohn et
al. (2003, 2004)]. In Shu et al. (2011), the method presented
here is used to reject six harmonics. Also applications where
the frequencies are not independent but harmonically related
will be considered (treating them as independent would
introduce conservatism and might not lead to a solution). The
tracking performance will also be investigated in greater
detail.
ACKNOWLEDGEMENTS
The authors wish to thank Sennheiser corporation for
providing the active noise control headset. Helpful
discussions with Dr. Hatem Foudhaili (Sennheiser) about
active noise control from an industrial perspective were also
greatly appreciated.
REFERENCES
Apkarian, P. and P. Gahinet. (1995). A convex
characterization of gain-scheduled H  controllers.
IEEE Transactions on Automatic Control, vol. 40, pp.
853-864.
7902
Ballesteros, P. and C. Bohn. (2011). A frequency-tunable
LPV controller for narrowband active noise and
vibration control. Proceedings of the American Control
Conference. Accepted for publication.
Bohn, C., Cortabarria, A., Härtel, V. and K. Kowalczyk.
(2003). Disturbance-observer-based active control of
engine-induced vibrations in automotive vehicles.
Proceedings of the SPIE’s 10th Annual International
Symposium on Smart Structures and Materials, Paper
No. 5049-68.
Bohn, C., Cortabarria, A., Härtel, V. and K. Kowalczyk.
(2004). Active control of engine-induced vibrations in
automotive vehicles using disturbance observer gain
scheduling. Control Engineering Practice, vol. 12, pp.
1029-1039.
Du, H. and X. Shi. (2002). Gain-scheduled H  control for
use in vibration suppression of system with harmonic
excitation. Proceedings of the American Control
Conference, pp. 4668-4669.
Du, H., Zhang, L. and X. Shi. (2003). LPV technique for the
rejection of sinusoidal disturbance with time-varying
frequency. IEE Proceedings on Control Theory and
Applications, vol. 150, pp. 132-138.
Feintuch, P. L., Bershad, N. J. and A. K. Lo. (1993). A
frequency-domain model for filtered LMS algorithms -
Stability analysis, design, and elimination of the training
mode. IEEE Transactions on Signal Processing, vol. 41,
pp. 1518-1531.
Francis, B. and W. Wonham. (1976). The internal model
principle of control theory. Automatica, vol. 12, pp.
457-465.
Hansen, B. (2001). Understanding active noise cancellation.
Spon, London.
Heins, W., Ballesteros, P. and C. Bohn. (2011). Gain-
scheduled state-feedback control for active cancellation
of multisine disturbances with time-varying frequencies.
Proceedings of the 10th MARDiH Conference on Active
Noise and Vibration Control Methods. Accepted for
publication.
Hong, J. H. and D. S. Bernstein. (1998). Bode integral
constraints, colocation, and spillover in active noise and
vibration control. IEEE Transactions on Control
Systems Technology, vol. 6, pp. 111-120.
Hostetter, G. H. (1980). Fourier-analysis using spectral
observers. Proceedings of the IEEE, vol. 68, pp. 284-
285.
Kinney, C. E. and R. A de Callafon. (2006a). Scheduling
control for periodic disturbance attenuation.
Proceedings of the American Control Conference, pp.
4788-4793.
Kinney, C. E. and R. A de Callafon. (2007). A comparison of
fixed point designs and time-varying observers for
scheduling repetitive controllers. Proceedings of the
46th IEEE Conference on Decision and Control, pp.
2844-2849.
Kinney, C. E. and R. A. de Callafon. (2006b). An adaptive
internal model-based controller for periodic disturbance
rejection. Proceedings of the 14th IFAC Symposium on
System Identification, pp. 273-278.
Kuo, S. M. and D. R. Morgan. (1996). Active noise control
systems: Algorithms and DSP implementations. Wiley,
New York.
Shu, X., Ballesteros, P. and C. Bohn. (2011). Active vibration
control for harmonic disturbances with time-varying
frequencies through LPV gain scheduling. Proceedings
of the 23rd Chinese Control and Decision Conference.
Accepted for publication.