Journal of Sound and Vibration (1998) 214(4), 784–791

Article No. sv981566

HYBRID CONTROL: SEPARATION IN DESIGN

R. L. C
Department of Mechanical Engineering and Materials Science, Duke University, Durham,
North Carolina 27708-0300, U.S.A.

D. S. B
University of Michigan, Department of Aerospace Engineering, Ann Arbor,
Michigan 48109-2118, U.S.A.
(Received 12 September 1997, and in final form 9 February 1998)

1. 
Active control of sound and vibration has gained significant interest in recent years due
to advances in digital signal processing hardware, the application of adaptive signal
processing in the control of sound and vibration, and the development of new transduction
devices for what have been frequently termed ‘‘adaptive’’, ‘‘smart’’, or ‘‘intelligent’’
structures. The application of adaptive control approaches, based largely upon the LMS
algorithm and its derivatives, has proceeded in parallel with efforts in the controls
community devoted to the design of fixed-gain, robust compensators [1]. Advantages of
hybrid (adaptive feedforward and feedback) control has been discussed in recent years
[2–4]; however, there has been very little effort devoted to the interpretation of the two
control strategies from a common terminology base.
Within the control literature, the ‘‘standard problem’’ is frequently used as the basis for
control system design and synthesis [5–6]. This standard problem essentially involves the
development of a state variable model, obtained from analysis or experimental system
identification, with a convenient structure. The block diagram describing the standard
problem has been termed the two-input, two-output (TITO) model. A schematic diagram
of this system is presented in Figure 1. As illustrated, there are two vector inputs, w(s)
and u(s), and two vector outputs, z(s) and y(s). The two vector inputs are the disturbance
and control respectively while the two vector outputs are the error and measurement
respectively. The generalized plant, G(s), is divided into separate matrix transfer functions
between each vector input and vector output. The performance path is defined between the
disturbance input, w(s), and the error output, z(s), and this path essentially defines the cost
function. The control path is defined between the control input, u(s), and the measured
output, y(s), and is used to implement the controller, K(s). The reference path defines the
relationship between the disturbance and the measured output and the secondary path
defines the relationship between the control and error output.
The generalized plant shown in Figure 1 contains the dynamics of the system to be
controlled as well as any additional frequency weighted filters that describe physical
processes or penalties imposed by the designer. For example, in the structural acoustic
control problem, measurements are obtained and control inputs are applied through the
dynamic structure (beam, plate, shell, etc.); however, a model of the fluid–structure
interaction must be constructed such that the controller is designed to minimize sound
radiation as opposed to vibration. This fully coupled model describes the generalized plant,
G(s). In general, the designer’s knowledge of the application physics is conveyed in the
model of the generalized plant.

0022–460X/98/290784 + 08 $30.00/0 7 1998 Academic Press

     785
G (s)
w (s) z (s)
Gzw (s) Gzu (s)
Exogenous Error
inputs output
u (s) Gyw (s) Gyu (s) y (s)
Control Measured
inputs output

K (s)

Figure 1. Block diagram of the standard problem.

Another convenience offered upon casting the system model in terms of the standard
problem results when the system response is assumed linear (a basic assumption for a large
class of acoustic and vibration control applications). Linear, time invariant (LTI) state
variable representations can be used to represent the system dynamics, providing an
efficient means of computing the system response to steady state or transient inputs.
Methods of designing fixed-gain, dynamic compensators for feedback control of such
systems are prevalent in the literature. The difficulty in developing such models most
frequently results when additional application physics must be incorporated into the plant
model to generate the error signals used in the cost functional. For example, in structural
acoustic control, one can develop an expression relating the surface velocity to the radiated
power. When outputs are expressed in units of power, spectral factorization must be
employed to develop a state variable model of such dynamics. Thus, the application of
these models required to develop the cost functional are limited to steady state predictions
in control system design. However, such techniques are commonly employed in the design
of feedback control as detailed by references [7–11].
The purpose and contribution of this work rests in the unified description of feedback,
feedforward, and hybrid control system architectures within the general framework of the
standard problem. The effects of feedforward, feedback, and hybrid control on the plant
dynamics are illustrated. The standard problem is modified to demonstrate the specialized
structure of the generalized plant when feedforward or hybrid control is applied.
Additionally, as demonstrated in this work, the feedback and feedforward control system
components are shown to be separable with respect to all disturbances which can be
measured directly.

2.   

2.1. Feedback control
As detailed in the introduction, the standard problem is structured as illustrated in
Figure 1, and the relationship between the inputs and outputs of the open-loop plant can
be expressed as

$ % $ %$ %
z(s) Gzw (s) Gzu (s) w(s)
= . (1)
y(s) Gyw (s) Gyu (s) u(s)

As indicated in equation (1) and illustrated in Figure 1, the system matrix G(s) is
partitioned according to [z/y] and [w/u]. The closed-loop response at the performance
786    
output as a function of an input disturbance can be obtained through a linear fractional
transformation,
z(s) = Tzw (s)w(s), (2)
where
Tzw (s) 0 [Gzw (s) + Gzu (s)K(s)[I − Gyu (s)K(s)]−1Gyw (s)], (3)
= [Gzw (s) + Gzu (s)K(s)S(s)Gyw (s)], (4)
and
S(s) 0 [I − Gyu (s)K(s)]−1, (5)
is the sensitivity function. The closed-loop response is thus dependent upon all transfer
function paths and the dynamic compensator.
As noted by Hong and Bernstein[12], a special case occurs when the control actuator
is colocated with the disturbance (i.e., Gzw (s) = Gzu (s) and Gyw (s) = Gyu (s)) or when the
performance (error) variable is the same as the measured variable (i.e., Gzw (s) = Gyw (s) and
Gzu (s) = Gyu (s)). In these cases, the closed-loop transfer function, Tzw (s), reduces to the
simplified expression
Tzw (s) = S(s)Gzw (s). (6)
Thus, if smax (S( jv)) Q 1 (where smax (S( jv)) is the maximum singular value of the
sensitivity function as a function of frequency), then smax (Tzw ( jv)) Q smax (Gzw ( jv)). This
special case is the multi-variable version of the classical, single input, single output
servo-control problem in which open loop-shaping is used for compensator design.
However, for the more general problem outlined in equation (4), a reduction in sensitivity
may not lead to a direct improvement of closed loop response, and thus closed loop
shaping must be applied through more general techniques.

2.2. Feedforward control

In feedforward control, the entries of G(s) can be described as illustrated in Figure 2.
The disturbance input to the structure can be measured directly (i.e., Gyw = I), and the
control signal has no influence on the measured output (i.e., Gyu = 0). Although the
feedforward control problem can be cast in the architecture of the standard problem, the
controller cannot be synthesized through H2 design approaches since the plant is
unobservable from the measured output, y(s) (i.e., the measured disturbance or reference
signal as denoted in adaptive feedforward control). However, if a feedforward

G (s)
w (s) z (s)
Gzw (s) Gzu (s)
Exogenous Error
inputs output
u (s) I 0 y (s)
Control Measured
inputs output

K (s)

Figure 2. Block diagram of the standard problem configured for feedforward control.
     787
compensator, K(s), is selected, the response between the disturbance and performance
variables is given as

Tzw (s) = Gzw (s) + Gzu (s)K(s). (7)

Thus, only the zeros of the system response can be modified since polynomial addition will
occur only in the numerator. As a result, any reduction in smax (Tzw ( jv)) is local to the
output z(s) since the original poles of Gzw (s) are not modified by the application of
feedforward control [13].

2.3. Hybrid control

Hybrid control, in this work, is defined as the combination of feedback and feedforward
control. The hybrid control problem is cast in terms of the standard problem in Figure 3.
The disturbance has been partitioned into a random input, wr (s), and a measured input,
wm (s). The measured output has also been partitioned into an output measured directly
from the plant response, yP (s), due to the inputs u(s) and w(s) and a direct measure of
the measured input, ywm (s). The generalized plant, G(s), has been partitioned appropriately
for the input and output variables. Notice also that the dynamic compensator has been
partitioned to reflect the portion associated with feedback control, KuyP (s) and the portion
associated with feedforward control, Kuwm (s). Through simple matrix algebra, one can show
that the controlled response between the disturbance and performance can be described
as

z(s) = [Gzw (s) + Gzu (s)KuyP (s)[I − GyPw (s)KuyP (s)]−1GyPu (s)]w(s)

+ [Gzu (s) + Gzu (s)KuyP (s)[I − GyPu (s)KuyP (s)]−1GyPu (s)]Kuwm (s)wm (s). (8)

Equation (8) can be partitioned so as to separate the disturbance due to the random input,
wr (s) and that due to the measured input, wm (s):

z(s) = [Gzwr (s) + Gzu (s)KuyP (s)[I − GyP u (s)KuyP (s)]−1GyP wr (s)]wr (s)

+ ([Gzwm (s) + Gzu (s)KuyP (s)[I − GyP u (s)KuyP (s)]−1GyP wm (s)]

+ [Gzu (s) + Gzu (s)KuyP (s)[I − GyP u (s)KuyP (s)]−1GyP u (s)]Kuwm (s))wm (s). (9)

G (s)
wr
w=
wm z
Gzwr (s) Gzwm (s) Gzu (s)
yp
y=
yw
u Gypwr (s) Gypwm (s) Gypu (s) m

0 I 0

Kuyp (s) Kuwm (s)

Figure 3. Block diagram of the standard problem configured for hybrid control.
788    
In adapative feedforward control applications, the performance output variables, and the
measured output variables are frequently different; however, if they are the same (i.e., when
z(s) = yP (s)) equation (9) can be simplified to
z(s) = ([I − Gzu (s)KuyP (s)]−1Gzwm (s) + [I − Gzu (s)KuyP (s)]−1Gzu (s)Kuwm (s))wm (s)
+ [Gzwr (s) + Gzu (s)Kuz (s)[I − Gzu (s)Kuz (s)]−1Gzwr (s)]wr (s). (10)
Through manipulation one can show that
z(s) = S(s)[Gzwm (s) + Gzu (s)Kuwm (s)]wm (s) + S(s)Gzwr (s)wr (s). (11)
Thus, if smax (S(jv)) Q 1, then a reduction in the response due to the random input
disturbance, wr ( jv), will result. However, the response due to the measured disturbance
input, wm (s), is controlled through a combination of zero placement (feedforward control)
and the magnitude reduction in the sensitivity (feedback control). For control of
harmonics, the zeros are readily modified to ‘‘cancel’’ the output due to the disturbance.

3.       

In this section one applies LQG theory to the hybrid system shown in Figure 3. One
demonstrates that the LQG controller can be decomposed into two components, namely,
a feedback controller and a feedforward controller. In addition, the LQG controller is
shown to satisfy a separation property in the sense that the feedback component of the
controller coincides precisely with the LQG controller that is obtained if the tonal
disturbance is absent.
Consider the following linear time-invariant (LTI) state equation:
ẋ(t) = Ax(t) + Bu(t) + D1r wr (t) + D1m wm (t), (12)
where wr (t) is a random disturbance and wm (t) is a disturbance that can be measured
directly. One can construct two outputs:
yP (t) = Cp x(t) + D2P wP (t), (13)
and
ywm (t) = wm (t) + Dms wms (t), (14)
where yP (t) is the output filtered by the plant to be controlled, and ywm (t) is the output
corresponding to the disturbance which can be measured directly. Sensor noise is added
to each output and is represented by D2P wP (t) and Dms wms (t) respectively. The disturbance
and sensor noise signals wr , wm , wP , and wms are mutually uncorrelated white noise inputs
with unit intensity. The performance or error output is represented as
z(t) = E1 x(t) + E2 u(t), (15)
in terms of the system states, x(t), and control input, u(t).
In the usual notation, the system equations can be expressed as
ẋ(t) = Ax(t) + Bu(t) + D1 w, y(t) = Cx(t) + D2 w(t), (16,17)

where

$ % $ %
0 0 D2P 0 CP
D1 = [D1r D1m 0 0], D2 = , C= , (18–20)
0 I 0 Dms 0
     789
and
K wr (t) L
G wm (t) G
w(t) = G G. (21)
wP (t)
G G
k wms (t) l

The resulting Riccati equations for the hybrid LQG problem can be expressed as
0 = (A − BR−1 T T −1 T −1 T −1 T
2 R12 ) P + P(A − BR2 R12 ) − PBR2 B P + R1 − R12 R2 R12 , (22)
0 = (A − V12 V−1 −1 T T −1 −1 T
2 C)Q + Q(A − V12 V2 C) − QC V2 CQ + V1 − V12 V2 V12 , (23)
where
R1 = E1T E1 , R12 = E1T E2 = 0, R2 = E2T E2 ,
V1 = D1 D1T , V12 = D1 D2T , V2 = D2 DT2 .
Note that there is no cross-weighting assumed in the performance measure. The state
variable description for the compensator can be expressed as
Ac = A + BCc − Bc C, Bc = (QC T + V12 )V−1
2 , Cc = −R−1 T T
2 (B P + R12 ). (24–26)
Next one notes that V1 , V12 , and V2 are given by
V1 = D1r D1rT + D1m D1m
T
= V1r + V1m , V12 = [0 D1m ], (27, 28)

$ %
T
D2P D2P 0
V2 = T , (29)
0 I + Dms Dms

where
V1r = D1r D1rT T
V1m = D1m D1m .
Now, assuming that Dms is zero,

$ %$ %
V−1
2P 0 CP
V12 V−1
2 C = [0 D1m ] = 0, (30)
0 I 0

and

$ %$ %
V−1
2P 0 0
V12 V−1 T
2 V12 = [0 D1m ] T = V1m . (31)
0 I D1m

Substracting equation (31) from equation (27), one obtains

V1 − V12 V−1 T
2 V12 = V1r . (32)

Additionally,

$ %$ %
V−1
2P 0 CP
CV−1 T
2 C = [CP 0] = CPT V2P CP . (33)
0 I 0
790    
Thus, equation (23) can be expressed as

0 = AQ + QA T − QCPT V−1
2P CP Q + V1r . (34)

However, this is exactly the Riccati equation that would result if one ignores wm throughout
the development and simply uses yP .
Making appropriate substitutions in equation (25), one can show that

Bc = [QCPT V−1
2p D1m ]. (35)

Thus, QCPT V−1

2p is the gain for yp , and D1m is the gain for ym . The final compensator can
be written as

ẋc (t) = Ac xc (t) + Bc y(t) = Ac xx (t) + QCPT V−1

2P yp (t) + D1m ym (t) (36)

and

u(t) = Cc xc (t). (37)

Note that if wm = 0 (i.e., the measured input disturbance is no longer applied to the
system), then ywm = 0. For this case,

ẋc (t) = Ac xc (t) + QCPT V−1

2P yP (t). (38)

This is identical to the LQG controller that would result if wm were set to zero from the
start and serves to demonstrate the independence of the feedback and feedforward design
for stochastic and measured disturbance inputs respectively. Thus, if a portion of the
disturbance can be measured and controlled with feedforward, or adaptive feedforward
control, this portion can be ignored during the design process required to develop a
compensator for control of stochastic disturbance inputs.

4. 
The hybrid (feedback and feedforward) control problem was cast in the format of the
standard problem described in the controls literature. Since both feedback and feedforward
control approaches are frequently applied to noise and vibration problems, a method of
considering the design objectives in a concurrent format was presented.
The hybrid measurement structure and problem was formulated to demonstrate that the
LQG compensator design for control of stochastic inputs is separable with respect to
inputs which can be measured directly. Thus, one can design a feedback controller in the
absence of measureable input disturbances if the objective is to combine both feedback
and feedforward or adaptive feedforward control for stochastic and measureable inputs
respectively. Once the feedback controller is designed for the stochastic inputs, the
feedforward controller required for the measurable input disturbances can be designed in
a subsequent formulation of the control problem. Realizing that the design of such
compensators is separable serves to simplify the hybrid design process.
Future work will be devoted to the application of such hybrid control system design to
problems in vibration and acoustics. Additionally, the hybrid design approach will be
explored to determine its affect on order reduction for dynamic compensation.

ACKNOWLEDGMENTS
This work was supported in part by the NSF under grant CMS 9501470 and AFOSR
under grant AFOSR FY9620-95-1-0019.
     791

1. K. Z, J. C. D and K. G 1996 Robust and Optimal Control. New Jersey: Prentice
Hall.
2. R. L. C 1995 ASME Journal of Dynamic Systems, Measurement, and Control 117, 232–240.
A hybrid autonomous control approach.
3. W. R. S, H. H. R and R. B 1996 Noise Control Engineering Journal
44, 11–21. A hybrid structural control approach for narrowband and impulsive disturbance
rejection.
4. W. T. B 1997 Journal of Sound and Vibration 205, 121–133. An adaptive feedback
approach to structural vibration suppression.
5. B. F 1986 A Course in Ha Control Theory. New York: Springer-Verlag.
6. S. B and C. B 1991 Linear Controller Design: Limits of Performance. Englewood
Cliffs, NJ: Prentice Hall.
7. N. K. G 1980 Journal of Guidance and Control 3, 529–535. Frequency shaped cost
functionals: extension of linear-quadratic–Gaussian design methods.
8. W. T. B, W. R. S and H. H. R 1991 Journal of the Acoustical Society
of America 90, 3202–3208. Active suppression of acoustic radiation from impulsively excited
structures.
9. W. T. B, F. S. H and H. H. R 1992 Journal of the Acoustical Society of
America 92, 1998–2005. Active structural acoustic control of broadband disturbances.
10. D. R. T and P. A. N 1995 Journal of the Acoustical Society of America 98,
2651–2662. Feedback control of sound radiation from a plate excited by a turbulent boundary
layer.
11. R. L. C and D. E. C 1998 (in press) in Journal of the Acoustical Society of America.
Multi-variable structural acoustic control with static compensation.
12. J. H and D. S. B 1996 Proceedings of the Conference of Control Applications
928–933, Dearborn, MI. Bode integral constraints and spillover in active noise and vibration
control. (also, IEEE Trans. Contr. Sys. Tech., to appear).
13. R. L. C, W. R. S and G. P. G 1998 Adaptive Structures: Dynamics and Control.
New York: John Wiley.