FEEDBACK SYSTEMS

113

Other alternatives exist for analyzing the stability of asymmetric systems. Walker (1970),
approaching the problem by looking for Lyapunov functions, was able to state several results
for the stability of Equation (4.18) in terms of a fourth matrix R. If there exists a symmetric
positive definite matrix R such that RA1
1 A3 is symmetric and positive definite (this is the
1
same as requiring A1 A3 to be symmetrizable), then the system is stable if the symmetric part
1
of RA1
1 A2 is positive semidefinite and asymptotically stable if the symmetric part of RA1 A2
is strictly positive definite. This result is slightly more general than the symmetrizable results
just stated in that it allows the equivalent symmetric systems to have gyroscopic forces.
In addition to these results, Walker (1970) showed that, if there exists a symmetric matrix
1
R such that RA1
1 A2 is skew-symmetric and RA1 A3 is symmetric, and such that R and
1
RA1 A3 have the same definiteness, then the system is stable but not asymptotically stable.
Another approach to the stability of Equation (4.18), not depending on symmetrizable
coefficients, has been given by Mingori (1970). He showed that, if the coefficient matrices
M1 D1 G1 H1 and K satisfy the commutivity conditions
HD1 M = MD1 H
HD1 G = GD1 H
HD1 K = KD1 H
then the stability of the system is determined by the matrix
Q = HD1 MD1 H GD1 H + K
This theory states that the system is stable, asymptotically stable, or unstable if the matrix Q
possesses nonnegative, positive, or at least one negative eigenvalue respectively. Although
the problem addressed is general, the restrictions are severe. For instance, this method cannot
be used for systems with semidefinite damping (D1 does not exist).
Other more complicated and more general stability conditions are due to Walker (1974)
and an extension of his work by Ahmadian and Inman (1986). The methods are developed
by using Lyapunov functions to derive stability and instability conditions on the basis of
the direct method. These are stated in terms of the symmetry and definiteness of certain
matrices consisting of various combinations of the matrices A1 1 A2 , and A3 . These conditions
offer a variety of relationships among the physical parameters of the system, which can aid
in designing a stable or asymptotically stable system.

4.10 FEEDBACK SYSTEMS

One of the major reasons for using feedback control is to stabilize the system response.
However, most structures are inherently stable to begin with, and control is applied to improve
performance. Unfortunately, the introduction of active control can effectively destroy the
symmetry and definiteness of the system, introducing the possibility of instability. Thus,
checking the stability of a system after a control is designed is an important step. A majority
of the work in control takes place in state space (first-order form). However, it is interesting to
treat the control problem specifically in mechanical or physical coordinates in order to take
advantage of the natural symmetries and definiteness in the system. Lin (1981) developed

114

STABILITY

a theory for closed-loop asymptotic stability for mechanical structures being controlled by
velocity and position feedback. The systems considered here have the form (see Section 2.3)
M q + A2 q + Kq = ff + f

(4.23)

where M = M T is positive definite, A2 = D + G is asymmetric, and K is symmetric. Here, the

vector f represents external disturbance forces (taken to be zero in this section) and the vector
ff represents the control force derived from the action of r force actuators represented by
(4.24)

f f = Bf u

where the r 1 vector u denotes the r inputs, one for each control device (actuator), and Bf
denotes the n r matrix of weighting factors (influence coefficients or actuator gains) with
structure determined by the actuator locations. In order to be able to feed back the position
and velocity, let y be an s 1 vector of sensor outputs denoted and defined by
y = Cp q + Cv q

(4.25)

Here, Cp and Cv are s n matrices of displacement and velocity influence coefficients

respectively, with structure determined by the sensor locations and where s is the number
of sensors. Equation (4.25) represents those coordinates that are measured as part of the
control system and is a mathematical model of the transducer and signal processing used to
measure the system response. The input vector u is chosen to be of the special form
u4t5 = Gf y = Gf Cp q Gf Cv q

(4.26)

where the r s matrix Gf consists of constant feedback gains. This form of control law is
called output feedback, because the input is proportional to the measured output or response, y.
In Equation (4.24) the matrix Bf reflects the location on the structure of any actuator or
device being used to supply the forces u. For instance, if an electromechanical or piezoelectric
actuator is attached to mass m1 in Figure 2.4, and if it supplies a force of the form F0 sin t,
the vector u reduces to the scalar u = F0 sin4t5 and the matrix Bf reduces to a vector
BfT = 61 07. Alternatively, the control force can be written as a column vector u, and Bf
can be written as a matrix




F0 sin t
1 0
and
Bf =
u=
0 0
0
If, on the other hand, there are two actuators, one attached to m1 supplying a force
F1 sin41 t5 and one at m2 supplying a force F2 sin42 t5, then the vector u becomes uT =
6F1 sin41 t5 F2 sin42 t57 and the matrix Bf becomes Bf = I, the 2 2 identity matrix.
Likewise, if the positions x1 and x2 are measured, the matrices in Equation (4.25) become
Cp = I and Cv = 0, the 2 2 matrix of zeros. If only the position x1 is measured and the
control force is applied to x2 , then


0
Bf =
0

0
1

and

1
Cp =
0

0
0

FEEDBACK SYSTEMS

115

Making the appropriate substitutions in the preceding equations and assuming no external
disturbance (i.e., f = 0) yields an equivalent homogeneous system, which includes the effect
of the controls. It has the form
M q + 4D + G + Bf Gf Cv 5q + 4K + Bf Gf Cp 5q = 0

(4.27)

For the sake of notation, define the matrices D = Bf Gf Cv and K = Bf Gf Cp . Note that, since
the number of actuators r is usually much smaller than the number of modeled degrees of
freedom n (the dimension of the system), the matrices K and D are usually singular. Since,
in general, D + D and K + K may not be symmetric or positive definite, it is desired to
establish constraints on any proposed control law to ensure the symmetry and definiteness
of the coefficient matrices and hence the stability of the system (see problem 4.11). These
constraints stem from requiring D + D and K + K to be symmetric positive definite. The
problem of interest in control theory is how to choose the matrix Gf so that the response
q has some desired property (performance and stability). Interest in this section focuses on
finding constraints on the elements of Gf so that the response q is asymptotically stable or
at least stable. The stability methods of this chapter can be applied to Equation (4.27) to
develop these constraints. Note that the matrices Bf Gf Cp and Bf Gf Cv are represented as the
matrices Kp and Kv respectively in Equation (2.17).
Collocated control refers to the case where the sensors are located at the same physical
location as the actuators. If the sensors or the actuators add no additional dynamics, then
collocated controllers provide improved stability of the closed-loop system. As seen above,
the closed-loop system coefficients D and K generally lose their symmetry for many
choices of Bf , Cf , Bv , and Cv . If, however, the gain matrices Gf and Gv are symmetric, and
if BfT = Cf and BvT = Cv , then the matrices D and K remain symmetric. The symmetry then
results in the possibility of choosing the gain matrices so that D and K remain positive
definite, ensuring closed-loop stability for stable open-loop systems (D and K positive
definite). Placing sensors and actuators at the same location causes BfT = Cf and BvT = Cv ,
so that collocated control enhances closed-loop stability. The controller design consists of
choosing gains Gf and Gv that are symmetric and positive definite (or at least semidefinite)
with collocated sensors and actuators to ensure a stable closed-loop response.

Example 4.10.1
Consider the two-degree-of-freedom system in figure 2.4, with a control force applied to m1 and a
measurement made of x2 so that the control system is not collocated. Then the input matrix, output
matrix, and symmetric control gain matrix are


1
Bf =
0


0
1
0

0
Cp =
0


1
1
0

g
G= 1
0

0
g2

Note that this is not collocated because BfT 6= Cf . The closed-loop system of Equation (4.27) becomes


m1
0

0
m2



 
x 1
c + c2
+ 1
x 2
c2

c2
c2



 
x 1
k + k2
+ 1
x 2
k2

k2 + g1
k2



  
x1
0
=
x2
0

116

STABILITY

This has an asymmetric displacement coefficient, implying the potential loss of stability. If, on the
other hand, a control force is applied to m1 and a measurement made of x1 , then the control system
is collocated and the input matrix, output matrix, and control gain matrix are


1
Bf =
0


0
1
0

1
Cp =
0


0
1
0

g
G= 1
0

0
g2

so that BfT = Cf and the closed-loop system of Equation (4.27) becomes



m1
0

0
m2



 
x 1
c + c2
+ 1
x 2
c2

c2
c2



 
x 1
k + k2 + g1
+ 1
x 2
k2

k2
k2



  
x1
0
=
0
x2

which is symmetric and stable for any choice of g1 such that k1 + k2 + g1 > 0.

The topic of control and Equation (4.27) is discussed in more detail in section 6.6 and
in Chapter 7. Historically, most of the theory developed in the literature for the control of
systems has been done using a state-space model of the structure. The next section considers
the stability of systems in the state variable coordinate system.

4.11

STABILITY IN STATE SPACE

In general, if none of the stability results just mentioned is applicable, the problem can be
cast in first-order form as given in Section 2.3. The system of Equation (4.18) then has
the form
x = Ax

(4.28)

where A is a 2n 2n state matrix and x is a 2n state vector. In this setting, it can easily be
shown that the system is asymptotically stable if all the eigenvalues of A have negative real
parts and is unstable if A has one or more eigenvalues with positive real parts.
The search for stability by finding a Lyapunov function in first-order form leads to the
Lyapunov equation
AT B + BA = C

(4.29)

where C is positive semidefinite and B is the symmetric, positive definite, unknown matrix
of the desired (scalar) Lyapunov function:
V 4x5 = xT Bx

(4.30)

Do not confuse the arbitrary matrices B and C used here with the B and C used for input
and output matrices. To see that V4x5 is, in fact, the desired Lyapunov function, note that
differentiation of Equation (4.30) yields
d
6V4x57 = x T Bx + xT Bx
dt

(4.31)