State Observers for Different Types of Systems

Tran Duc Long

Ngày 13 tháng 6 năm 2023

2.1 Introduction
This chapter begins with an introduction to the theories of state observers using the state-space de-
scription of dynamical systems. Consider the state-space representation of a linear system as described
by

ẋ(t) = Ax(t) + Bu(t) (2.1)

y(t) = Cx(t) (2.2)

Where x(t) ∈ Rn , u(t) ∈ Rm , and y(t) ∈ Rp are the state vector, the control input vector, and the
measurement output vector, respectively. Matrices A ∈ Rn×n , B ∈ Rn×m , and C ∈ Rp×n are known
constants.
The state variable representation of systems has some advantages over the more conventional
transfer function representation [77]. The state vector x(t) contains enough information to completely
summarize the past behavior of the system, and the future behavior is governed by the ordinary
differential equation:

ẋ(t) = Ax(t) + Bu(t)

The properties of the system are determined by the constant matrices A, B, and C. Thus the study
of the system can be carried out in the field of matrix theory which is well developed and has many
notational and conceptual advantages.
When faced with the problem of controlling a system, some schemes must be devised to choose
the input vector u(t) so that the system behaves in an acceptable manner. Since the state vector
x(t) contains all the essential information about the system, it is reasonable to base the choice of
u(t) solely on some combinations of x(t). In other words, u(t) is determined by a relation of the form
u(t) = −Kx(t) where K ∈ Rm×n is a constant matrix.
As described in Chapter 1, in most control situations, the state vector is not available for direct
measurement. This means that it is not possible to directly evaluate the function Kx(t). Luenberger
[77] showed how the available system inputs and outputs may be used to construct an estimate of the
system state vector. The device which reconstructs the state vector is called an observer. The observer
itself, as a time-invariant linear system, is driven by the inputs and outputs of the system it observes.
It was further illustrated in [77] how the time constant of the observer can be chosen arbitrarily
Over the past three to four decades, there has been an increasing percentage of control systems
literature written in the state variable point of view. Modern optimal control laws ( [7,73] ) almost
invariably have control commands based ultimately upon the state variables. In addition, useful design
tools such as pole placement are most easily implemented in a state feedback framework. As previously
mentioned, a complete state vector is rarely available for use in state feedback, so it is necessary for the
design to be modified. A useful approach is to use state estimates in place of actual state variables for
use in the control law, and these estimates will come from an observer. Prior to discussing the role of
observers in reconstructing an incomplete or inaccessible state vector of a linear system, first consider
a linear feedback system with a completely accessible state vector. If the state x(t) of the open-loop
system is completely available, then a linear feedback control law is of the form:

u(t) = −Kx(t). (2.3)

1
 The matrix K ∈ Rm×n is designed to pass a linear combination of state variables into the system
inputs for use in a control law. When the input is applied to (2.1), the closed-loop system is then

ẋ(t) = (A − BK)x(t). (2.4)

The dynamic response of the state vector in (2.4) can be asymptotically driven to the desired zero
vector if a feedback gain matrix K can be found, such that the matrix (A − BK) has eigenvalues in
the left-hand side of the complex s-plane.
This is a paragraph of text that will appear between the two figures.

Fig. 2.1: Closed-loop State Feedback Control

A conceptual diagram of a full-state feedback control scheme is shown in Figure 2.1 where the control
law u(t) = −Kx(t) is computed assuming full knowledge of the state vector x(t). This system
regulates itself against unknown disturbances affecting the output, by inputting the control law to
hold the output steady. It can be readily modified to regulate itself around some desired reference
signal for a tracking application.
As demonstrated, the dynamic response of a controllable system can be altered through the
implementation of a linear state feedback controller (2.3). Such an application requires that system
parameters and the current state x(t) of the system be completely available. In the event that the
states are unavailable, a state observer is required to reconstruct or estimate the states that are
required as inputs to the controller. The observer should simulate the actual system accurately, to
provide an estimation of the states required for feedback. For the observer to be of any practical use,
the error between the model and the real system must be asymptotically zero.

Fig. 2.2: Closed-loop Observer for State Feedback

Figure 2.2 illustrates the role of the observer in reconstructing the states of the system so that it
may be used for feedback into the controller. The concept of a compensator is introduced, which is
simply an observer and controller working together to provide the feedback path.

2
 In this chapter, various types of state observers will be discussed in detail, to facilitate a platform
for the development of linear functional observers and to broaden the class of linear systems to include
unknown inputs, time-delay, non-linear and descriptor systems.

2.2 Full-Order State Observers for Linear Systems

As the name suggests, the full-order state observer estimates all the states of the system. Luenberger
[78] defined this observer and introduced the conditions which must be satisfied in order for the observer
to exist. It follows that a system with n state variables requires an observer of order n for full state
estimation. Consider the linear time-invariant system of (2.1) & (2.2), where

ẋ(t) = Ax(t) + Bu(t) (2.5)

y(t) = Cx(t). (2.6)

For this system, it is now assumed that all state variables contained in x(t) cannot be measured
directly and that an observer to reconstruct the state variables must be formed. To reconstruct all the
state variables, x̂(t) is introduced (where x̂(t) denotes an approximation of the state x(t) as opposed
to the actual state). Define an observer with the following dynamics

˙
x̂(t) = Ax̂(t) + Bu(t) + L(y(t) − C x̂(t)) (2.7)

where L ∈ Rn×p is the observer gain matrix, to be determined as stated in Theorem 2.1 to follow.
Figure 2.3 shows the input and output of a full-order state observer and also how the estimated state
x̂(t) is used for feedback control purpose. This form of state observer uses the system input, u(t), and
output, y(t), to estimate the state vector, x(t). The first two terms are a model of the system dynamics
while the third term is dependent on the difference between the actual output and the expected output
based on the current estimated state, and thus represents the mismatch between the system model
and the actual system.

Fig. 2.3: Schematic of a Full-Order State Observer

Let the observer state estimation error be defined as follows:

△
e(t) = x(t) − x̂(t) (2.8)

Then by taking the derivative of (2.8) and substituting (2.5), (2.6) and (2.7), the error dynamics
can be obtained as follows

3
 ˙
ė(t) = ẋ(t) − x̂(t)
= (Ax(t) + Bu(t)) − (Ax̂(t) + Bu(t) + L(y(t) − C x̂(t)))
= (Ax(t) + Bu(t)) − (Ax̂(t) + Bu(t) + L(Cx(t) − C x̂(t)))
= (A − LC)x(t) − (A − LC)x̂(t).

Incorporating the definition in (2.8), this collapses to the following differential equation

ė(t) = (A − LC)e(t) (2.9)

This differential equation can be readily solved, the solution being an exponential function of the
form

e(t) = e(A−LC)t e(t0 ). (2.10)

Clearly, for the estimation error to approach zero, the eigenvalues of (A − LC) must have negative
real parts, or equivalently (A − LC) must be Hurwitz. Since the system is time-invariant, the system
parameters A and C are constant. The matrix L must therefore be chosen to satisfy those conditions
via the method of pole positioning. Following the Principle of Duality, presented in [90], and based
upon linear state feedback results, this concept is formalized in the following Theorem.

Theorem 2.1.(Theorem (1.16) [90])

Corresponding to the real matrices A and C, there is a real matrix L such that the set of eigenvalues
of (A − LC) can be arbitrarily assigned (subject to complex eigenvalues occurring in conjugate pairs)
if and only if the pair (C, A) is completely observable.

This important result defines the requirement for a full-order state observer to exist. It highlights
that if the original system is unobservable then it is not possible to construct a full-order state observer.
Assuming that the system is observable, the question of how the location of the poles will affect
the performance of the observer must be addressed. Since the purpose of the observer is to provide the
system states to the controller, the observer must provide an estimated signal to the controller that
tends toward the actual states rapidly and accurately. Otherwise, the controller will yield an incorrect
control signal u(t) = −K x̂(t). The positioning of poles in (A − LC) determines several performance
characteristics of the observer, of which the most important is the observer transient response. As the
real component of the poles becomes increasingly negative (move further left in the complex s-plane),
the estimated states tend toward the actual states more rapidly. On the other hand, this also results
in the elements of L becoming larger, thus amplifying sensor noise and resulting in a less accurate
control signal.
In the Luenberger observer design, one defining characteristic of the state estimation process is that
of the error trajectory. Let there be a second-order system with two state variables x1 (t) and x2 (t).
Subsequently, the error functions can be defined simply as e1 (t) = x1 (t)−x̂1 (t) and e2 (t) = x2 (t)−x̂2 (t),
where x̂1 and x̂2 represent the estimated values for x1 and x2 , respectively. Figure 2.4 illustrates the
error trajectory that may result when plotting e2 (t) against e1 (t). As can be seen, starting at an
arbitrary point in the error space, the plot may follow a spiral trajectory approaching the origin
asymptotically.
Consider the state feedback using the reconstructed states with the following control law

u(t) = −K x̂(t) (2.11)

Assuming an appropriate choice for matrix L, the estimation error e(t) will be zero in the steady
state. The initial transient response of the full-order state observer will be investigated by considering
the full-order state observer presented in Figure 2.3. Of particular interest when analyzing the transient
response, is whether the introduction of the observer will degrade system stability.

4
 Fig. 2.4: Typical Luenberger Observer Error Trajectory

The closed-loop system is completed through the interconnection of the controller (2.11) and observer
(2.7) to the open-loop system (2.5) & (2.6). It is described by the following composite system
    
ẋ(t) A −BK x(t)
˙ = (2.12)
x̂(t) LC (A − LC − BK) x̂(t)
 
  x(t)
y(t) = C 0 (2.13)
x̂(t)
From the observer definition (2.7), the definition of the reconstruction error (2.8), and prescribed
control law (2.11), ẋ(t) and ė(t) can be derived as follows:

ẋ(t) = Ax(t) + Bu(t) = Ax(t) − BK x̂(t)

= Ax(t) − BK(x(t) − e(t))
= (A − BK)x(t) + BKe(t), (2.14)
ė(t) = ẋ(t) − x̂˙ = Ax(t) − Ax̂(t) − L(Cx(t) − C x̂(t))
= (A − LC)e(t). (2.15)
In terms of the error between the model and real system, this results in
    
ẋ(t) (A − BK) BK x(t)
= (2.16)
ė(t) 0 (A − LC) e(t)
To investigate system stability, consider the eigenvalues of the above system by analyzing the
characteristic polynomial
  
(A − BK) BK
det λI2n −
0 (A − LC)



= det λIn − (A − BK) × det λIn − (A − LC) (2.17)

5
 This leads to a result on eigenvalue assignability and asymptotic stability of the overall closed-loop
control system represented by (2.12).
Theorem 2.2. (Theorem (1.17) [90])
If the linear system is completely controllable and completely observable, there exist gain matrices K
and L such that the 2n eigenvalues of the system matrix of the closed-loop system can be arbitrarily
assigned, in particular to positions in the left-half of the complex s-plane.

From (2.17), it is clear that the eigenvalues of the combined observer-controller are the union of the
eigenvalues of two matrices (A − BK) and (A − LC). In the control literature, this important prop-
erty is sometimes referred to as the separation principle. What this implies is that in the controller
design process, the state feedback controller for the system can first be designed, assuming that all
state variables are available for feedback. The controller should be designed to satisfy some system
specifications. Then a state observer with a faster dynamics than the closed-loop control system can be
designed. Finally, the estimated state variables are used to generate an estimate of the control signal.
This signal is then fed back to the original system. This design process preserved the stability of the
overall closed-loop system.

 The proof of the separation

  principle
 can also be derived from (2.12) by pre- and post-multiplying
A −BK I 0
by , i.e.,
LC (A − LC − BK) I −I

     
I 0 A −BK I 0 (A − BK) BK
eig( ) = eig( ).
I −I LC (A − LC − BK) I −I 0 (A − LC)
Note that the introduction of an observer results in an impaired transient response to the closed-
loop system, as a result of the non-zero reconstruction error. If designed correctly, this degradation in
performance diminishes exponentially along with the error e(t), thus demonstrating the application of
the state reconstructor. This form of observer is simple in its implementation. According to Theorem
2.1, as long as the system is fully observable, there are no restrictions on the pole positions of the
observer. There is also no added limit on the generality of the observer as there is no requirement
for the system to be given in a specific canonical form. However, the observer does suffer from the
fact that it reconstructs the whole state vector, which may not necessarily be required. This inherent
redundancy will be addressed in the design of reduced-order state observers in Section 2.3.

2.2.1 Design Algorithm

1. Check if the pair (C, A) is observable. If not: stop, as a state observer with arbitrary
poles does not exist.
Otherwise, go to the next step.
2. Choose the desired pole locations for the observer.
3. Use any pole-placement algorithm (for example, the place algorithm in MATLAB) to find
L so that (A − LC) has the desired poles for the observer.

2.2.2 A Numerical Example

Consider a 9th -order system with the system matrices A, B and C are as given below.
   
1 0 −2 2 0 0 −8 −6 −2 1

 0 −1 3 18 41 −10 −19 −28 4 

2
 

 1 2 3 4 4 0 −8 −8 4 

3
   

 1 0 0 1 0 0 0 −1 1 

4
  1 0 0 0 0 0 0 0 0
A=
 2 1 0 0 −2 3 2 0 1 
 , B = 5 andC =  0
  1 0 0 0 0 0 0 0 .

 −3 1 0 0 1 0 −1 0 0 

6
  0 0 1 0 0 0 0 0 0

 −1 0 −1 1 3 0 −1 1 1 

7
 
 0 0 1 2 1 0 4 −5 1  8
1 1 0 1 0 1 −2 −3 0 9

6
 The observability matrix has full rank and therefore the pair (C, A) is observable. If the poles for
the observer be chosen to be {−1, −2, −3, −4, −5, −6, −7, −8, −9} then using the place algorithm in
MATLAB matrix L is obtained as follows:

 
16.4513 1.1666 0.7849

 11.1405 6.5434 27.5687 


 4.2599 1.0679 18.0053 


 2.1650 0.3792 12.0118 

L=
 6.7402 2.2080 7.7137 .

 8.2183 1.8452 7.7330 

 8.4099 1.0155 1.2267 
 4.2763 0.5504 2.9643 
0.4285 0.4886 8.0546

Since matrix (A − LC) has eigenvalues in the left-half of the complex s-plane, the observer esti-
mation error goes to zero asymptotically. The full-order observer that estimates the entire state vector is

˙
x̂(t) = Ax̂(t) + Bu(t) + L(y(t) − C x̂(t)).