Sie sind auf Seite 1von 22

EE5101 Linear Systems

Project Report
Semester I, 2010/11





Name : Rakesh Viyak Gunasekaran
Matriculation No : A0076997B
E-Mail ID : rakeshviyak@nus.edu.sg






Linear Systems Project Report A0076997B
Page | 2


Table of Contents


1. ABSTRACT ........................................................................................................................................ 3
2. INTRODUCTION ................................................................................................................................... 4
3. STATE-SPACE REPRESENTATION AND SYSTEM ANALYSIS AND DYNAMICS ................................... 5
System analysis ................................................................................................................................... 6
Controllability and observability. .................................................................................................... 6
Stability analysis .................................................................................................................................. 6
4. LINEAR QUADRATIC OPTIMAL CONTROL ............................................................................................ 9
LQR control design .............................................................................................................................. 9
Simulation of Plant Dynamics ............................................................................................................. 9
5. STATE FEEDBACK AND OUTPUT FEEDBACK CONTROL...................................................................... 13
Full-Order observer ........................................................................................................................... 13
Output-Feedback control .................................................................................................................. 13
Simulation for the State and Error Dynamics ................................................................................... 14
6. SERVO CONTROL ............................................................................................................................... 18
7. Conclusions ....................................................................................................................................... 21
8. Reference .......................................................................................................................................... 22










Linear Systems Project Report A0076997B
Page | 3



1. ABSTRACT

Consider the plant shown in the figure below, where
1 2
2
1 2
( ) , ( ) ,
2 5
G s G s
s as a s
= =
+ +

My Matriculation Number is A0076697
So a = 6997/1000 = 6.997.








We need to design some controllers based on the desired performance using the
techniques like pole placement, Linear Quadratic Optimal Design, observer and Servo
control to remove the unwanted rejection of disturbance in the system, here in this case
the disturbance is step input.
Based on the state-space description, we consider many approaches in designing the
control laws capable of stabilizing the given system. First, we present a LQR state-
feedback control law that optimally steers the nonzero initial states to the origin, while
still ensure the stability of closed-loop system.
In practice, it is usual that a state-feedback control law cannot be physically
implemented due to the unobservability of system states. As a consequence, an observer
is typically employed to estimate the unmeasured states. We will also introduce a well-
known full-order observer, and then based on this observer and the pole-placement
method. We present an output-feedback control law capable of stabilizing the given
system with prescribed control performance.
Finally, we introduce two control laws with unity output feedback, of which one is based
on the pole-placement method and the other is not, to solve the problem of servo
control for the given system in the presence of disturbance which is ubiquitous in
practice. The control problem is thus practically meaningful.
) ( s Y
) ( s D
) ( s U
1
1
2 2
4 . 1 a as s
+ +
) (
1
s G ) (
2
s G
Linear Systems Project Report A0076997B
Page | 4

In each section, apart from presenting theory, we do many simulations to show the
effectiveness of the proposed control laws. It is shown that the simulation results in all
sections correspond with the theory presented.

2. INTRODUCTION

To understand the basic concepts of controller designing using various methods
like pole placement method, in order to obtain the desired performance from the given
system.


[1] To find a state-space model for the plant and to analyze and simulate the plants
dynamic characteristics;
[2] To design a state feedback controller using the LQR method and to simulate the
designed system for showing the state responses to non-zero initial state with zero
external inputs.
[3] To design a state observer using the pole placement method and to simulate the
resultant observer-based feedback control system and to study the closed-loop
control performance;
[4] To design a controller such that the output error response goes to zero at the steady-
state when reference signal and the disturbance input d(t) are both of step type and
to show the output time responses to reference and disturbance respectively.

System analysis and control synthesis are two component parts of the system
control theory. For linear systems, these parts are well developed and investigated.
However, many of, if not almost of, the tools and frameworks in system analysis and
control synthesis of other complex systems, such as nonlinear systems, networked
systems, and hybrid systems, are inspired by those of linear systems. As such, a well
understanding about linear control systems equips us a strong background to do
research deeply in the system control theory. In this project, we aim at investigating, as
a way to build up the understanding of system control theory, several methods and
approaches in analyzing and controlling a given plant described as a linear system.
As an unstable system is usually of no use in practice, the stability property is the most
important dynamic characteristic of a system. As a consequence, typical duties in
system analysis and control design are to investigate the stability property of the
system and to stabilize the unstable system. However, to guarantee that the control
signal can render the system to be stable, other characteristics such as controllability
and observability are of interest. By representing the given plant in the state-space
model, we are able to analyze such important dynamic characteristics of the plant.
Linear Systems Project Report A0076997B
Page | 5


3. STATE-SPACE REPRESENTATION AND SYSTEM ANALYSIS AND
DYNAMICS

The open loop transfer function for the output Y(s) is written as,
2
2
1 1
( ) ( ) ( )
4.284 9.3636
1 1
( ) ( )
( 4.284 9.3636)
Y s U s D s
s s s
U s D s
s s s s
(
= +
(
+ +

= +
+ +


The given system is described in terms of block diagram representation. For more
convenient in analyzing and controlling the system, in this section, we will find the systems
state-space model, for which many well developed tools and frameworks in system analysis and
control synthesis have been built.
Let y1 be the output of the G1(S) and y is the output of the G2(S).
The input to the second system G2(S) is given by y1 + d.
The variable d is the denoted as disturbance.
Let the state variables , be the output after the block , and . From the
given figure, we derive the state-space model of the plant as follow
1 =
2 =
3=
(1)
Where u and y are system input and output, respectively, and d is disturbance.
By denoting , we can rewrite (1) to a matrix form

(2)
where
Linear Systems Project Report A0076997B
Page | 6



Given,


a = 6997/1000 = 6.997
Then,
A =

System analysis
Controllability and observability.
We achieve the controllability matrix of the system as follow
U = [ B AB A
2
B]=
Det (U) = -1 (3)
We also have the observability matrix of the system as follow
O = [C CA CA
2
]
T
=

Det (O) = 8 (4)
From (3) and (4), we can conclude that matrices U and O are non-singular, therefore
rank(U)=rank(O)=3, in other word U is a full row rank matrix and O is a full column rank matrix.
Hence, the given system is controllable and observable.
Stability analysis

We can find that the transfer function of the open-loop system G(S) = G1(S) G2(S), has a
pole on the imagine axis. Therefore, the open-loop system is not BIBO stable.
The eigen value of the matrix A = [0, -10.71, -3.2609] (5)
The matrix A has two eigen values with negative real part and one eigen value with zero real
part. Consequently, the open-loop system is not asymptotically stable in sense of Lyapunov.
Linear Systems Project Report A0076997B
Page | 7

Simulation of plant dynamics
The simulation results are shown below. Figure 1 and 2 show the step response and
impulse response of the plant output, while Figure 3 shows the dynamics of system states. It can
be seen from Figure 1 that the system output diverges, or the system is not BIBO stable. Figure 2
shows that system is not asymptotically stable as its states do not converge to the origin after
being affected by impulse signal.

Figure 1: Step Response
Linear Systems Project Report A0076997B
Page | 8


Figure 2: Impulse Response

Figure 3: State Dynamics of the Plant
Linear Systems Project Report A0076997B
Page | 9

4. LINEAR QUADRATIC OPTIMAL CONTROL
In the previous section, we have analyzed several dynamic characteristics of the system.
It is shown that the system is not stable in the sense of BIBO stability and asymptotically stable
in sense of Lyapunov. As such, we need to design control laws to stabilize the system. We first
investigate a simple control design framework, namely LQR control design, to produce a state-
feedback control law that stabilizes the system optimally.
LQR control design
In the case of zero external inputs, the plant is described by the following standard
linear state-space model

(6)
Our objective is to render the non-zero initial states of the system to the origin optimally. In the
other words, our objective is to design a control law u such that the closed-loop system is
asymptotically stable, while a given performance index is minimized. Such an objective is
guaranteed by solving the LQR problem, in which the cost index takes the linear quadratic form
as follows
(7)
where Q is a symmetric positive-semidefinite matrix, and R is a symmetric positive-definite
matrix, accounting for the expenditure of the energy of the control signals.
The matrices Q and R determine the relative importance of the error and the expenditure of the
control energy, and more precisely, they determine the relative importance of keeping x and u
near zero. As the matrix Q is symmetric positive-semidefinite, there exists the matrix H such
that
It is proved that for the above LQR problem, if the pair (A, B) is controllable and the pair (A, H) is
observable, then the optimal control law has the form of state feedback as follows
(8)
where P is a symmetric positive-definite matrix satisfying the Algebraic Riccati Equation:
(9)
Moreover, the optimal control law (8) guarantees that the closed-loop system is asymptotically
stable. Therefore, our objective is well achieved.
Simulation of Plant Dynamics

We have proved that (A, B) is controllable. For numerical convenience, we choose Q =
diag (q, q, q), and thus, H = diag . To ensure (A, H) is observable, we must
choose q>0. Inversely, if q>0 then rank (H) = 3, an thus, the matrix H is of full column rank. As a
Linear Systems Project Report A0076997B
Page | 10

consequence, the matrix is also full column rank and (A, H) is observable.
Therefore, the LQR problem in this case is solvable.
As it is the relative weight between Q and R that is importance, and u is scalar, we choose R = 1
for numerical convenience.
With a = 6.997, we have
A =


In this project, we choose the initial states
x1(0) = -4
x2(0) = 8
x3(0) = 6
in all the simulations.
We first choose q = 100. From (8) and (9), the respective optimal control gain is
K = [10.0000 5.9391 3.5477]
Simulation results are shown in Figures 4-5 which illustrates that the closed-loop system is
asymptotically stable, and all the signals in the system tend to zero.
Linear Systems Project Report A0076997B
Page | 11


Figure 4: System State Dynamics using LQR with Q=100

Figure 5: Control Input and System Output with Q=100
Linear Systems Project Report A0076997B
Page | 12

We can interpret these observations as follows. From cost function (7), we can see that the
larger the matrix Q (and thus q) is, the faster convergence behaviours of system states x are well
achieved. However, from Algebraic Riccati Equation (9) and control law (8), the larger the
matrix Q is, the larger the matrix P is and thus, the larger control law u is.
In practice, every control signal must be reasonably small to be physically implementable.
Attaching special importance to practical aspect, we choose the design parameters q = 1 to
ensure the implement ability of the control law u in practice, while still guarantees the
reasonably fast convergence behaviours of system states.
For Q=0.01 I K = [0.1000 0.0401 0.0032]
For Q=0.1 I K = [0.3162 0.1278 0.0127]
For Q=1 I K = [1.0000 0.4137 0.0652]
For Q=10 I K = [3.1623 1.4198 0.4515]
For Q=100 I K = [10.0000 5.9391 3.5477]

Figure 6: Impulse Response for different values of Q

Linear Systems Project Report A0076997B
Page | 13


5. STATE FEEDBACK AND OUTPUT FEEDBACK CONTROL

In the previous section, we have utilized the state-feedback control law to stabilize the
given plant with zero external inputs. In practice, however, the system state x is not fully
measurable, leading to the state-feedback control laws as in (8) are not implementable.
Furthermore, an undesired behaviour of the closed-loop system using a state-feedback
controller is the sensitivity to uncertainties. An observer capable of estimating the unmeasured
system states by using measurable system inputs and outputs is thus naturally needed.
Full-Order observer
A well-know state estimator (observer), introduced by Luenberger in 1964, is full-order
observer.
For the above observable plant (6), the full-order observer is as follows
( ) x C y L Bu x A x + + =


(10)
Where x denotes the estimate of x and the matrix L is to be designed. From (6) and (10), the
dynamics of the estimation error x x e = is given by

( )
e LC A
x C y L x x A
x C y L Bu x A Bu Ax x x e
) (
) ( ) (
] [ ) (
=
=
+ + + = =



(11)
As the pair (A, C) is observable, we can choose the matrix L such that the eigenvalues of (A-LC)
are arbitrary, particularly, we choose L such that (A-LC) is stable. This together with (11) leads
to that 0 e when t . Hence, the state estimate x tracts the system state x asymptotically.
Output-Feedback control
Replacing the system state x in the state-feedback control law by the estimated state x , we
obtain the output-feedback control law
( ) x C y L Bu x A x + + =


x K r u =
(12)
Thus we have

Br BKe x BK A
e x K r B Ax
x K r B Ax x
+ + =
+ =
+ =
) (
)] ( [
) (

Linear Systems Project Report A0076997B
Page | 14


This together with (11) leads to that the closed-loop system can be expressed as follow
r
B
e
x
LC A
BK BK A
e
x
(

+
(

=
(

0

0


(13)
Let
(

= I
LC A
BK BK A
0

, then det (sI- I) = det (sI-(A-BK) )det(sI-(A-LC)). As (A-BK) and (A-LC)
are both stable, we conclude that the closed-loop system is asymptotically stable in the sense of
Lyapunov. By that equality, we also see that an output-feedback control law stabilizing a linear
system can be designed by combining two separated frameworks: designing a state-feedback
control law to stabilize the linear system and designing a state observer to estimate the system
states. Replacing the system states in the state-feedback control law by the estimated state, we
obtain a stabilizing output-feedback control law.


Simulation for the State and Error Dynamics
we have obtained the optimal control gain (Q=1) K = [1.0000 0.4137 0.0651]. The
respective poles of closed-loop system, i.e., the roots of characteristic polynomial can be
obtained by equation det (sI-(A-BK)) = 0
Roots are -0.0286, -10.7638, -3.2468
Note that, we must ensure the convergence behaviours of the estimation errors are relatively
fast in comparison to the convergence behaviours of the system states. As such, we should
choose the observer gain L such that the observer poles are 3-5 times larger than the control
poles. This is easy to be obtained since (A, C) is observable. We choose the observer poles as
p1 = -10 p2 = -1+j p3 = -1-j

Using the matlab function we can obtain the value of L

L = [-0.9970 7.4595 -59.5084]
T
Simulation results are shown in Figures 7-9. It can be seen that the convergence behaviours of
all signals in the systems are well achieved.

Linear Systems Project Report A0076997B
Page | 15


Figure 7: State Space dynamics of the observer


Figure 8: Error Dynamics of the observer

Linear Systems Project Report A0076997B
Page | 16


Figure 9: Control input and the output
Now we investigate the effects of observer poles on state estimation errors and closed-loop
control performance. To guarantee the control performance is reasonably satisfactory, we
choose all observer poles are p = -15, then the L is obtained as,
L=[ 15.5030 103.0585 -297.0734]
T
Simulation results in this case are shown in Figures 10-11. It can be seen from Figures 10-11
that the convergence behaviours of state estimation errors as well as system states are much
faster than the previous case. These simulation results are very plausible because when the
observer poles increase, the dynamics of estimation errors are very fast, leading to the
convergence behaviour of system states are also very fast.
However, the control signal is also very large as seen from Figure 13. In practice, every control
signal must be reasonably small to be physically implemented. As such, we should choose the
observer poles that are not very large to ensure the implement ability of control signal. Clearly,
the control performances, derived by using the above output feedback framework, are arbitrary
due to the arbitrary placement of control poles and observer poles.
Linear Systems Project Report A0076997B
Page | 17


Figure 10: State dynamics of the system for faster poles

Figure 11: Error dynamics for faster poles
Linear Systems Project Report A0076997B
Page | 18

6. SERVO CONTROL

We have considered the control design problem for the given plant without the presence
of disturbances. In practice, however, the presence of disturbances is ubiquitous. As such,
considering the effects of disturbances in the control design is an inevitable task. The effects of
disturbance are typically eliminated by solving the servo control problem which includes
asymptotic tracking and disturbance rejection. In this section we utilize output-feedback control
law in solving the servo control design to deal with the existence of immeasurable states.

k
+
1
C
x =Ax +Bu
.
k
2
s
_
+
r
E
w
y
_

Figure 12: Multi Variable Integral control
Consider an m*m plant:
w
x Ax Bu B w
y Cx
= + +
=

The error is defined as
e = r-y
Suppose that the disturbance (w) and reference (r) are both of step type. To achieve zero steady
state error, like SISO integral control, we introduce one integrator to each channel;
0
( ) ( )
t
v t e t dt =
}

Then,
( ) ( ) ( ) ( ) v t e t r y t r Cx t = = = ,
Using above equations form an augmented system:
Linear Systems Project Report A0076997B
Page | 19

0 0
0 0 0
w
x A x B B
u w r
v C v I
| | | || | | | | | | |
= + + +
| | | | | |

\ . \ .\ . \ . \ . \ .

| |
0
x
y C
v
(
=
(


The controllability of the augmented system can be checked by the matrix:

It is certainly that the matrix is non-singular. In other word, the augmented system is controllable. It
can be stabilized by the control law
| |
1 2
x
u k k
v
(
=
(


The feedback gain K can be determined by pole placement method or LQR method. We choose LQR
method to obtain the feedback gain K.
As a result, the resultant feedback system is
1 2
0
0 0
w
x A Bk Bk x B
w r
v C v I
| | | || | | | | |
= + +
| | | | |

\ . \ .\ . \ . \ .

| |
0
x
y C
v
(
=
(


Once the feedback system is stable, each signal in the system in response to step inputs will be
constant in the steady state, and so is v (t). It follows that
( ) ( ) 0 e t v t = =
Thus, zero steady state error can be achieved.
We define Q=I and R=1.Then we obtain the value of the feedback gain K by simulating on the
MatLab, which is shown below.
K = [0.3765 4.8117 12.6272 -1.0000]
First, we set the reference signal r (t) as a step type signal while keep the disturbance signal
d(t)=0. The output time response to the step reference input is shown as below.
Linear Systems Project Report A0076997B
Page | 20

0 5 10 15 20 25 30 35 40
0
0.2
0.4
0.6
0.8
1
1.2
1.4
Output when refrence signal r(t) is step type and disturbance input d(t)=0

Figure 13: Output time responses to step reference input without disturbance input
From the figure above, with a step reference input and no disturbance input, we can find that the
output time response goes up a little bit and then goes down again. The output time response
converges to its input value after reasonable time.
Next, the disturbance signal d(t) is selected as a step type signal while keep the reference signal
r(t)=0. The output time response to the step disturbance input is shown below.
0 5 10 15 20 25 30 35 40
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Output when refrence signal r(t) is 0 and disturbance input d(t) is step type

Figure 14: Output time responses to step disturbance input without reference input
Linear Systems Project Report A0076997B
Page | 21

From the figure above, due to the step disturbance input and no reference input, we can see that
the output time response goes up from 0 to nearly 4.5 and the goes down below zero. But the
output time response converges to its stable value zero after some reasonable time around 20 secs.


7. Conclusions

The project let us analyse the system dynamics and make the knowledge into
practice. In the first chapter, we have built the state space model for the given open-loop
plant. And analyse the state space characteristics such as controllability, observability
and stability. The open-loop system given is observable and controllable, but unstable
.In the second chapter, by following the steps to design a state feedback controller, the
LQR controller has been designed and the unstable given system become stable by using
feedback controller. After that we keep R (or Q) unchanged and vary Q (or R) to see the
effect of Q (or R) to the system performance. If we assign a large weight age to
important states, the magnitude of those states will be reduced. Thus assigning a large
value to state produced less overshoot in the output. But, when assigning a large value
to the R caused the controller to useless control signal. This resulted in a slower
response time for the output. In the following chapter, we design an observer, and vary
the poles to see which kinds of poles perform better. From the analysis, we can draw the
conclusion that, when the observer poles were slower than the closed loop poles of the
system, the estimated states converged slowly to the actual ones and so the state
estimation error took a longer time to settle to zero. But in contrary, the performance
improved and matched that of the actual states when the observer was designed using
fast poles.
The final chapter, we have designed a multivariable integral controller by servo
control and the LQR method to obtain the feedback gain. The output time responses to
reference and disturbance are shown respectively. It is obviously that the output
becomes the same as the reference input after some reasonable time and the response
to the disturbance input goes up at the beginning and converges down to zero after a
certain time. The strengths and weaknesses of the multivariable integral control
method are discussed as well. Last but not the least, this Project helped me gain a better
understanding of the theory which I learned from class and let me put the theory into
practice.

Linear Systems Project Report A0076997B
Page | 22


8. Reference

[1] Linear System Theory and Design by C.T.Chen. 3rd Edition, Published by Oxford University Press
1999.
[2] Wang, Q.G., Linear System Part II, Notes of National University of Singapore, pp.1-392, 2007
[3] Prof Wang Qing-Guo, MATLAB Manual 2001

Das könnte Ihnen auch gefallen