You are on page 1of 12

# H Control

1
1.1

## Formulation of the H Problem

Uncertainty Review

## Consider the plant model with feedback as shown in Figure 1.

( s )

Error

z
External w Inputs
P( s)

External Outputs
y Measured

u Control Signals

K ( s)

Variables

Figure 1: Standard Presentation of Model with Uncertainty In this case (s) represents plant parameter uncertainty, P (s) represents the plant nominal transfer function and K(s) represents feedback control. 1.1.1 Sensitivity Reduction

The rst mission of the design is to make the system insensitive to external disturbances. This is equivalent to make z as independent of w as possible. 1

If we ignore the plant perturbation (s), then the Model simplies to that shown in Figure 2.

External w Inputs

P( s)

External Outputs
y Measured

u Control
Signals

K ( s)

Variables

Figure 2: Standard Presentation of Model without Plant perturbation (s) Suppose P (s) can be partitioned as follows

P11 (s) P12 (s) P (s) = P21 (s) P22 (s) so that z = P11 w + P12 u
[

y = P21 w + P22 u
]

then with the feedback law u = K(s)y we can eliminate u and y z = P11 + P12 K (I P22 K)1 P21 w (1)

= F1 (P, K)w To minimize the error z due to the external inputs w, we must minimize the function Fl (P, K). 1.1.2 Mixed Performance and Robustness Objective

The following set of characteristics are possible: We want to achieve good disturbance rejection from external signals in the low-frequency region. This can be achieved by making the sensitivity S = (I + P K)1 small as 0. Make the closed loop transfer function small at high frequencies limit excitation by noise. This can be achieved by making T = I S = I (I + P K)1 small as . 2

Guard against instability from parameter variations. This is achieved by minimizing K(I + P K)1 . We can then formulate the H problem as the minimization of the function

F1 (P, K) =

W1 S W3 (I S)

## Solution of the H Problem

It is possible to formulate the problem in many ways. In the literature a difference is made between the 1-block, 2-block and the 4-block formulations

2.1
2.1.1

Glover-Doyle Algorithm
Formulation

The Glover-Doyle algorithm is the classic formulation on which the Matlab Robust Control Toolbox concentrates. This toolbox solves the basic mixed performance and robustness objective. This algorithm solves a family of stabilising controllers such that Fl (P, K) Our search is to nd the lowest value of for which the above equation has a solution. One possibility is to start with the LQG solution and then to reduce it using a binary search. The plant equations in state space form is x = Ax + B1 w + B2 u z = C1 x + D11 w + D12 u y = C2 x + D21 w + D22 u and can be represented in the packed matrix form A B1 B2 C D G(s) = 1 11 D12 C2 D21 D22 3

The following assumptions must be satised to ensure a solution: The pair (A, B2 ) must be stabilizable and the pair (C2 , A) detectable. With the dimensions of dim x = n, dim w = m1 , dim u = m2 , dim z = p1 and dim y = p2 , then the Rank D12 = m2 and Rank D21 = p2 to ensure that they controllers are proper and the transfer function from w to y is non-zero at high frequencies (i.e. all-pass). A jI B2 Rank = n + m2 for all frequencies. C1 D12 A jI B1 Rank = n + p2 for all frequencies. C2 D21 D11 = 0 and D22 = 0 will simplify the equations and implies that the transfer functions from u to y and from w to z rolls o at high frequency. So our simplied problem is A B1 B2 G(s) = C1 0 D12 C2 D21 0 2.1.2 Solution

The solution of this problem requires the solving of two Ricatti equations, one for the controller and one for the observer. The control law is given by u = Kc x and the state estimator equation by x = Ax + B2 u + B1 w + Z Ke (y y ) where
T w = 2 B1 X x 2 T y = C2 x + D21 B1 X x

The controller gain is Kc as for the LQG case, and the estimator gain is Z Ke instead of Ke as for the LQG case, with
T T T Kc = D12 (B2 X + D12 C1 ), D12 = (D12 D12 )1 T T T Ke = (Y C2 + B1 D21 )D21 , D21 = (D21 D21 )1

and

Z = (I 2 Y X )1

The terms X and Y are solutions to the controller and estimator Ricatti equations X
T T T A B2 D12 D12 C1 2 B1 B1 B2 D12 B2 = Ric T T C1 C1 (A B2 D12 D12 C1

## T T with B1 = B1 (I D21 D21 D21 ) and C1 = B1 (I D12 D12 D12 ).

We do not carry out these calculations by hand the tools supplied by the Matlab Robust Control Toolbox does just that.

Properties of H Controllers
The stabilising feedback law u2 (s) = K(s)y2 (s) minimizes the norm of the closed loop transfer function Tyu = G11 (s) + G12 (s)[I K(s)G22 (s)]1 K(s)G21 (s) The problems we can solve is Optimal H2 control: min||Tyu ||2 Optimal H control: min||Tyu || Standard H control: min (||Tyu || 1) The H cost function Tyu is all-pass i.e. (Tyu ) = 1 for all values of . The H optimal controller (use hinfopt.m in Matlab) for an n-state augmented plant have at most n 1 states. 5

## The following important properties for H controllers exist:

The H sub-optimal controller (use hinf.m or the newer hinfsyn.m in Matlab) for an n-state augmented plant have exactly n states. In the weighted mixed sensitivity problem formulation, the H controller always cancels the stable poles of the plant with its transmission zeroes. In the weighted mixed sensitivity problem formulation, the unstable poles of the plant inside the specied bandwidth will be shifted to its mirror image once a H or H2 feedback loop is closed. The implications are that this technique allows very precise frequency-domain loop shaping via suitable weighting strategies. If you augment the plant with frequency dependent weights W1 to W3 , then the Matlab script hinf or newer hinfsyn or mixsyn will nd a controller that shapes the signals to the inverse of these weights, if it exists. The Matlab function augw.m forms the augmented plant

G(s) =

W1 W1 P 0 W2 0 W3 P I P

4
4.1

Examples
Example 1

## Consider the case of the double integrator G(s) = 1 s2

This plant violates the rules for a solution (poles on imaginary axis). Now we must set up the equations carefully. The equation set, in state space form, with the addition of a disturbance term representing uncertainty d, is x1 = d + u x2 = x1 with the regulated output (note the inclusion of the control signal to bound it) given by x2 z= u with the measurement equation y = x2 + n The noise term n may include measurement errors or unmodelled highfrequency dynamics we also need it to ensure the rank condition of D21 is met. Our set of equations are

0 0 A= , 1 0

1 B1 = , 0

1 B2 = , 0

0 1 , C1 = 0 0
[

C2 =

0 1

D11

0 0 = , 0 0

D12

0 = , 1

D21 =

0 1 ,

D22 = 0

In standard format and H format, the block diagrams of the double integrator is shown in Figure 3.

u+

d +

1 s

x1

1 s
y

x2

+ +

n
d +

1 s

x1

1 s

x2 +

+
G( s )

K ( s)

## (b) H block format

Figure 3: Double integrator example expressed into (a) Standard format and (b) H format Collecting the equations in packed matrix form
=

0 1 0

0 0 1

1 0 0

0 0 0

## 0 0 0 0 ... ... ... ... 0 1 0 1

. . . . . . . . . . . . . . .

1 0 0 1 ... 0

The solution to this problem by computer (pre-shifting the poles at the origin 8

[

1.59 1.08 ,

1.08 Ke = , 1.59

K(s) =

4.2

Example 2
s1 s+1

## Consider the following plant P (s) =

This need quite agressive control to stabilize. Lets focus on the sensitivity S and choose a weight 0.1 (s + 100) W1 = 100 s + 1 1 with bode plot (remember as W1 S 1, therefore S will track W1 )
Bode Diagram 20

Magnitude (dB)

20

W1(s)

40

60 0

Phase (deg)

45

90 10
4

10

10

10

Choose a moderate weight W2 = 0.1 and augment the plant with s W1 W2 W3 G = = = = = zpk(s); P = (s - 1)/(s + 1); % Plant is all-pass with zero in RHP 0.1*(s + 100)/(100 * s + 1); % Control S 0.1; % Moderate control on u []; % Ignore T augw(P, W1, W2, W3); % Augment the plant 9

The H -controller can be found using [K,CL,GAM] = hinfsyn(G); % or [K,CL,GAM] = mixsyn(G,W1,W2,[]); giving the controller, closed loop and = 0.1844 as K = 0.0001 TW1 S TW2 R (s + 1)(s 438900) (s + 0.01)(s + 69.53) 0.0009999 (s + 100)(s + 69.53)(s + 1)(s + 0.01) = (s + 0.01)(s + 1)(s + 1.876)(s + 23.77) 0.000009999(s + 1)2 (s 438900) = (s + 1)(s + 1.876)(s + 23.77)

How well did the H controller achieved the objectives? Generate the singular values using L = K*P; % Form loopgain S = inv(1+L); % Form S T = 1-S; % and T sigma(S,k,GAM/W1,k-.,T,r,GAM*P/W2,r-.) legend(S = 1/(1+L),GAM/W1, T=L/(1+L),GAM*P/W2,2) gives the following singular value plot
Singular Values 60 S = 1/(1+L) GAM/W1 T=L/(1+L) GAM*P/W2

40

20

40

60

80 4 10

10

10

10

10

10

4.3

Example 3

## Consider the following plant P (s) = 1 (s + 1)(s + 2)

Now choose weights to make the bandwidth about 3 rad/s and the sensitivity S as low as -40 dB at low frequencies. At the same time make the transmission T capable of robustly tolerating uncertainties of about 20 dB. Suitable weights would be (choose Ms = Mt = 1.5) 0.67(s + 4.5) s/Ms + s = s + s s s + 0.003 s + t /Mt 100(s + 2) W3 (s) = = s + 300 t s + t W1 (s) = Using hinfsyn we obtain = 1.1973 and the controller K= 110918138.86(s + 300)(s + 2)(s + 1) 3.01(s + 2)(s + 1) 2 + 3636s + 6495000) (s + 0.003)(s + 1701)(s s + 0.003

with matching
Singular Values 20 0 20 40 Singular Values (dB) 60 80 100 120 140 160 180 4 10 S = 1/(1+L) GAM/W1 T=L/(1+L) GAM*T/W3

10

10

10

The controller operates as a lead-type of controller, cancels the plant poles and uses the pole in W1 as its new controller pole. The match in S is very good but the rst-order pole makes the match in T quite bad. 11

By stiening the frequency requirement on T by using W3 = (W3 )2 as weight, reducing the overshoot to M = 1.2, we will arrive at = 1.72 and K 13.51(s + 2)(s + 1) (s + 0.003)(s + 7.09)
Singular Values 10 S = 1/(1+L) GAM/W1 T=L/(1+L) GAM*T/W3

20

30

40

50

60

70 3 10

10

10