Nonlinear Control

Chapter 8: Nonlinear Observers

In this chapter we review design of observers for nonlinear
systems. First a review on linear observer designs are given,
and the optimal Kalman filter formulation is introduce. Then
the use of this design for linearized system is given, and it is
generalized to extended Kalman filter with variable observer
gain. Then design on nonlinear observer for known and
uncertain system is elaborated.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 Welcome
To Your Prospect Skills
On Nonlinear System Analysis and
Nonlinear Controller Design . . .

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 Get more global exposure
Self confidence is the first key

About ARAS
ARAS Research group originated in 1997 and is proud of its 22+ years of brilliant background, and its contributions to
the advancement of academic education and research in the field of Dynamical System Analysis and Control in the
robotics application.ARAS are well represented by the industrial engineers, researchers, and scientific figures graduated
from this group, and numerous industrial and R&D projects being conducted in this group. The main asset of our
research group is its human resources devoted all their time and effort to the advancement of science and technology.
One of our main objectives is to use these potentials to extend our educational and industrial collaborations at both
national and international levels. In order to accomplish that, our mission is to enhance the breadth and enrich the
quality of our education and research in a dynamic environment.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 4
Contents
Local Observers
1 Motivation, review of linear observers, nonlinear system, fixed
observer gain, optimal observer (LQE), example.

Extended Kalman Filter (EKF)

2 Time-varing observer gain, linearization, differential Riccati
equation, assumptions, theorems, example.

Global Observers
3 Nonlinear but known systems, nonlinear Lipschitz systems,
extension of linear observer design.

High-Gain Observer
4
Motivating example; 2nd order system, nonlinear observer,
rejecting uncertainty, peaking phenomenon. Observer
formulation.

In this chapter we review design of observers for nonlinear systems. First a review on linear observer designs are given, and the optimal Kalman
filter formulation is introduce. Then the use of this design for linearized system is given, and it is generalized to extended Kalman filter with
variable observer gain. Then design on nonlinear observer for known and uncertain system is elaborated.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 5
Nonlinear Observers
• Motivation
 In many controller designs
• We need state-feedback
– Full state information is needed
– All state measurements are costly
• Could we use output-feedback
– Estimate the states from the outputs
– Observers are dynamical systems to do that
• In linear systems, we have separation theorem
– Design converging observer separately from stabilizing state-
feedback controller
– Stability of the closed loop output-controller is preserved.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 6
Nonlinear Observers
• Motivation
 Review observer design in linear systems
𝑥ሶ = 𝐴𝑥 + 𝐵𝑢, 𝑦 = 𝐶𝑥
• Full-state observer:
𝑥ොሶ = 𝐴𝑥ො + 𝐵𝑢 + 𝐻 𝑦 − 𝐶 𝑥ො
– If the pair (𝐴, 𝐶) is detectable
– 𝐻 can be designed such that the estimation error dynamics 𝑥෤ = 𝑥 − 𝑥ො
is stable
𝑥෤ሶ = 𝑥ሶ − 𝑥ොሶ = 𝐴 𝑥 − 𝑥ො − 𝐻 𝐶𝑥 − 𝐶 𝑥ො
𝑥෤ሶ = 𝐴 − 𝐻𝐶 𝑥෤
– For stability: design 𝐻 ∋ 𝐴 − 𝐻𝐶 is Hurwitz.
• This can be done on pole placement
• Or optimal LQE design
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 7
Nonlinear Observers
• Motivation
 For nonlinear system
𝑥ሶ = 𝑓(𝑥, 𝑢), 𝑦 = ℎ(𝑥)
• Design full-state observer:
𝑥ොሶ = 𝑓(𝑥,ො 𝑢) + 𝐻 ⋅ 𝑦 − ℎ(𝑥)
ො
– Where 𝐻(⋅) could be a constant or time-varying matrix.
– The term 𝑓(𝑥,
ො 𝑢) comes from model prediction
– The term 𝐻 ⋅ 𝑦 − ℎ(𝑥)
ො is a correcting term from observation model
• Two different approach could be used in observer design
– Linearization based methods
• Local Observers and EKF (both valid near equilibrium point)
– Global model based observers
• Model without uncertainty and with uncertainty

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 8
Nonlinear Observers
• Local Observers
 Consider the nonlinear system
𝑥ሶ = 𝑓 𝑥, 𝑢 , 𝑦 = ℎ 𝑥 (11.1)
• With full-state observer
𝑥ොሶ = 𝑓(𝑥,
ො 𝑢) + 𝐻 ⋅ 𝑦 − ℎ(𝑥)
ො 11.2
• The estimation error 𝑥෤ = 𝑥 − 𝑥ො satisfies:
𝑥෤ሶ = 𝑓 𝑥, 𝑢 − 𝑓(𝑥,
ො 𝑢) + 𝐻 ℎ(𝑥) − ℎ(𝑥) ො (11.3)
• The estimation error dynamics has an eq. point @ 𝑥 ෤ = 0.
• Design Objective:
– Design 𝐻 such that this eq. point is exponentially stable.
– Local solution for small 𝑥(0)
෤ .

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 9
Nonlinear Observers
• Local Observer
 Linearize (11.3) @ 𝑥෤ = 0:

• Design of a fixed 𝐻 to stabilize this TV system is impossible!

– Assume that ∃ 𝑥𝑠𝑠 ∈ 𝑅𝑛 , 𝑢𝑠𝑠 ∈ 𝑅𝑚 such that (11.1) has an eq. point:

– Assume that given 𝜀 > 0, ∃ 𝛿1 , 𝛿2 > 0, such that

– Ensure that x(t) satisfies:

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 10
Local Observer
 Design of a fixed 𝐻 (Cont.)
• Assume that the pair (𝐴, 𝐶) is detectable:

• Design H such that (𝐴 − 𝐻𝐶) is Hurwitz.

• Proof: One may show that:

– Where 𝐿1 is a Lipschitz constant of 𝜕𝑓/𝜕𝑥.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 11
Local Observer
• Proof (Cont.)
– Similarly

– Where 𝐿2 is a Lipschitz constant of 𝜕ℎ/𝜕𝑥.

• The estimation error dynamics is:

– Where:

– For some positive constants 𝑘1 , 𝑘2 .

• Lyapunov Candidate: 𝑉 = 𝑥෤ 𝑇 𝑃𝑥,
෤ where 𝑃 is pos. def. solution of

– Then for 11.5:

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 12
Local Observer
• Proof (Cont.)
– In which 𝑐4 = 2 𝑃 . Thus

– Which shows that for small 𝑥(0)

෤ , 𝜀 and 𝛿2 the estimation error
tends to zero as 𝑡 → ∞.
– To make sure 𝜀 is small enough, choose 𝛿1 , 𝛿2 sufficiently small. □

• Optimal Observer
 Linear-Quadratic Estimator (LQE)
• Consider LQE design for uncertain system
𝑥ሶ = 𝐴𝑥 + 𝐵𝑢 + 𝑤, 𝑦 = 𝐶𝑥 + 𝑣
– In which 𝑤 is the model uncertainty, while 𝑄 = Cov 𝑤 ,
– And 𝑣 denotes the observation noise, while 𝑅 = Cov 𝑣 .
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 13
Local Observer
• Full-state LQE observer:
𝑥ොሶ = 𝐴𝑥ො + 𝐵𝑢 + 𝐻 𝑦 − 𝐶 𝑥ො
– LQE optimal Gain:
𝐻 = 𝑃𝐶 𝑇 𝑅−1 (11.6)
– Where 𝑃 satisfies the Riccati Equation:

𝐴𝑃 + 𝑃𝐴𝑇 − 𝑃𝐶 𝑇 𝑅−1 𝐶𝑃 + 𝑄 = 0 (11.7)

– The constant matrices 𝑃, 𝑄, 𝑅 are symmetric and pos.def.

• Remark:
∞
– The LQE will minimize 𝐽 = 𝐸 ‫׬‬0 𝑥෤ 𝑇 𝑡 ⋅ 𝑥෤ 𝑡 𝑑𝑡
– 𝑃 and 𝐻 are constant in this design.
– Stabilizing the estimation error dynamics close to origin.
– See EKF formulation for time-varying optimal gain 𝐻(𝑡).

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 14
Local Observer
 Example
• Consider the system:

• where

– Check boundedness of 𝑥(𝑡). Take 𝑉1 (𝑥) = 𝑥 𝑇 𝑃1 𝑥,

3 1
– where 𝑃1 = is the solution of 𝑃1 𝐴1 + 𝐴1𝑇 𝑃1 = −𝐼. Then for 𝑥12 ≤ 2
1 1

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 15
Local Observer
 Example (Cont.)
• One may show that
min 𝑉1 𝑥 = 2
𝑥12 = 2

– Consider the set Ω = 𝑉1 𝑥 ≤ 2 ⊂ 𝑥12 ≤ 2 .

– Inside this set we have:

– Since the ball 𝐵𝑟 = 𝑥 ≤ 0.8081 is inside Ω.

– Hence 𝑉ሶ is negative on the boundary 𝜕Ω.
– Then Ω is positively invariant → The system trajectories are UUB.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 16
Local Observer
 Example (Cont.)
• Design LQE to estimate 𝑥(𝑡) for 𝑥 0 ∈ Ω.
• Consider 𝑄 = 𝐼, 𝑅 = 1;
– The solution to Riccati eq. 𝐴𝑃 + 𝑃𝐴𝑇 − 𝑃𝐶 𝑇 𝑅−1 𝐶𝑃 + 𝑄 = 0 is:
0.8105 −0.1715
𝑃= .
−0.1715 0.3284
– The LQE estimator gain is:
0.8105
𝐻 = 𝑃𝐶 𝑇 𝑅−1 = .
−0.1715
– Simulate the true system and the observer dynamics together:
𝑥ሶ = 𝐴1 𝑥 + 𝐵1 [0.25𝑥12 𝑥2 + 0.2 sin 2𝑡]
𝑥ොሶ = 𝐴1 𝑥ො + 𝐵1 𝑢ො + 𝐻 𝑦 − 𝐶1 𝑥ො , 𝑢ො = [0.25𝑥ො12 𝑥ො2 + 0.2 sin 2𝑡]

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 17
Local Observer
 Example (Cont.)
• Solve these equations numerically with:
– Initial conditions: 𝑥1 0 = 1, 𝑥2 0 = −1; 𝑥ො1 (0) = 𝑥ො2 0 = 0

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 18
Local Observer
 Example (Cont.)

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 19
Contents
Local Observers
1 Motivation, review of linear observers, nonlinear system, fixed
observer gain, optimal observer (LQE), example.

Extended Kalman Filter (EKF)

2 Time-varing observer gain, linearization, differential Riccati
equation, assumptions, theorems, example.

Global Observers
3 Nonlinear but known systems, nonlinear Lipschitz systems,
extension of linear observer design.

High-Gain Observer
4
Motivating example; 2nd order system, nonlinear observer,
rejecting uncertainty, peaking phenomenon. Observer
formulation.

In this chapter we review design of observers for nonlinear systems. First a review on linear observer designs are given, and the optimal Kalman
filter formulation is introduce. Then the use of this design for linearized system is given, and it is generalized to extended Kalman filter with
variable observer gain. Then design on nonlinear observer for known and uncertain system is elaborated.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 20
Local Observer
• Extended Kalman Filter
 Consider
• The nonlinear system (11.1) and the observer (11.2)
• With time-varying observer gain 𝐻(𝑡)
• The estimation error 𝑥෤ = 𝑥 − 𝑥ො satisfies:
𝑥෤ሶ = 𝑓 𝑥, 𝑢 − 𝑓(𝑥,
ො 𝑢) + 𝐻(𝑡) ℎ(𝑥) − ℎ(𝑥)
ො (11.8)
• Use Taylor series about 𝑥 ෤=0

– Where

– and

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 21
Extended Kalman Filter
 Kalman Filter Gain:

• Where 𝑃(𝑡) satisfies Differential Riccati Equation

– The constant matrices 𝑃0 , 𝑄, 𝑅 are symmetric and pos.def.

𝑄 = Cov 𝑤 , 𝑅 = Cov(𝑣) where 𝑤, 𝑣 are the noise of plant and
measurement respectively.
– Note the Riccati eq. (11.12) and the observer (11.2) shall be solved
simultaneously since 𝐻 𝑡 , 𝑃 𝑡 are time varying
– And 𝐴 𝑡 , 𝐶(𝑡) depends on 𝑥(𝑡)
ො .

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 22
Extended Kalman Filter

• Sketch of Proof:
– From Assumption 11.1 There exist 𝑃−1 𝑡 ∋ m (11.15)
– Consider Lyapunov candidate: 𝑉 = 𝑥෤ 𝑇 𝑃−1 𝑥෤
– and note 𝑃ሶ −1 = −𝑃−1 𝑃𝑃
ሶ −1

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 23
Extended Kalman Filter
• Sketch of Proof: (Cont.)
– Substitute 𝑃ሶ from Riccati eq. (11.12)

– Where 𝑃−1 𝑄𝑃−1 is uniformly pos.def in 𝑡.

– In view of (11.15) 𝐶 𝑇 𝑅−1 𝐶 is pos. semidefinite. Hence the sum is
uniformly pos.def. in t. Then

– For some positive constants 𝑐1 , 𝑐2 .

– Consequently,

1
– Where 𝑟 is the positive roots of − 𝑐1 + 𝑐2 𝑘1 + 𝑐2 𝑘2 𝑦 2 .
2
– Hence the origin is exponentially stable. □

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 24
Extended Kalman Filter
 Example
• Consider the same system:

• where

• Design EKF to estimate 𝑥(𝑡) for 𝑥 0 ∈ Ω.

• Consider 𝑄 = 𝑃 0 = 𝐼, 𝑅 = 1;
– The differential Riccati eq. (11.12):
𝑃ሶ = 𝐴𝑃 + 𝑃𝐴𝑇 − 𝑃𝐶 𝑇 𝑅−1 𝐶𝑃 + 𝑄, 𝑃 0 =𝐼
– where,

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 25
Extended Kalman Filter
 Example (Cont.)
𝑝11 𝑝12
– Take 𝑃 = 𝑝
12 𝑝22
– Then the EKF is defined by solving the five simultaneous equation:
– Simulate the true system and the EKF dynamics all together:
𝑥ሶ = 𝐴1 𝑥 + 𝐵1 [0.25𝑥12 𝑥2 + 0.2 sin 2𝑡]
𝑥ොሶ = 𝐴1 𝑥ො + 𝐵1 𝑢ො + 𝐻(𝑡) 𝑦 − 𝐶1 𝑥ො , 𝑢ො = [0.25𝑥ො12 𝑥ො2 + 0.2 sin 2𝑡]
𝑃ሶ = 𝐴(𝑡)𝑃(𝑡) + 𝑃(𝑡)𝐴 𝑡 𝑇
− 𝑃(𝑡)𝐶 𝑇 𝑅−1 𝐶𝑃(𝑡) + 𝑄,

• Simulate these equations numerically with:

– Initial conditions: 𝑥1 0 = 1; 𝑥2 0 = −1, 𝑥ො1 0 = 𝑥ො2 0 = 0, 𝑃 0 = 𝐼.
– Check the results.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 26
Extended Kalman Filter

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 27
Extended Kalman Filter
% EKF Example
% EKF Simulation function xp=ekf1(t,x)
x0=[1; -1; 0; 0; 1; 0; 1];
% State definition
tspan=[0 5];
% True states
[t,y]=ode45(@ekf1,tspan, x0, ...
odeset('OutputFcn','odeplot','outputSel',[3:4],'Max xt1=x(1); xt2=x(2); xt=[xt1;xt2];
Step',1e-1)); % Estimated states
subplot(2,1,1) xh1=x(3); xh2=x(4); xh=[xh1;xh2];
plot(t,y(:,1),'-b',t,y(:,3),'-.m',t,y(:,2),'--k', % P states
t,y(:,4),':r'),grid
title('True and Estimated States') p11=x(5); p12=x(6); p22=x(7); P=[p11 p12; p12
p22];
xlabel('Time (sec)')
ylabel('States') % System Matrices
legend('x_1','xh_1','x_2','xh_2') A=[0 1; -1 -2];
subplot(2,2,3) B=[0; 1];
plot(t,y(:,1)-y(:,3),'-b',t,y(:,2)-y(:,4),'- C=[1 0];
.k'),grid ut=0.25*xt1^2*xt2+0.2*sin(2*t);
title('EKF Estimation Errors') uh=0.25*xh1^2*xh2+0.2*sin(2*t);
xlabel('Time (sec)')
% Design parameters
ylabel('State Errors')
Q=eye(2,2); R=1;
legend('e_1','e_2')
subplot(2,2,4) % Total dynamics
plot(t,y(:,5),'-b',t,y(:,6),'--m',t,y(:,7),'- xtp=A*xt+B*ut;
.k'),grid y=C*xt;
title('EKF P Convergence') At=[0 1; -1+0.5*xh1*xh2 -2+0.25*xh1^2];
xlabel('Time (sec)') Ph=At*P+P*At'+Q-P*C'*inv(R)*C*P;
ylabel('P_i s') H=P*C'*inv(R);
legend('p_{11}','p_{12}','p_{22}')
xhp=A*xh+B*uh+H*(C*xt-C*xh);
set(findall(figure(1),'type','line'),'linewidth',2)
xp=[xtp;xhp;Ph(1,1);Ph(1,2);Ph(2,2)];
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 28
Contents
Local Observers
1 Motivation, review of linear observers, nonlinear system, fixed
observer gain, optimal observer (LQE), example.

Extended Kalman Filter (EKF)

2 Time-varing observer gain, linearization, differential Riccati
equation, assumptions, theorems, example.

Global Observers
3 Nonlinear but known systems, nonlinear Lipschitz systems,
extension of linear observer design.

High-Gain Observer
4
Motivating example; 2nd order system, nonlinear observer,
rejecting uncertainty, peaking phenomenon. Observer
formulation.

In this chapter we review design of observers for nonlinear systems. First a review on linear observer designs are given, and the optimal Kalman
filter formulation is introduce. Then the use of this design for linearized system is given, and it is generalized to extended Kalman filter with
variable observer gain. Then design on nonlinear observer for known and uncertain system is elaborated.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 29
Nonlinear Observers
• Global Observers
 Consider a nonlinear but known system
• Consider system in the form:

– where, 𝑥 ∈ 𝑅𝑛 , 𝑢 ∈ 𝑅𝑚 , 𝑦 ∈ 𝑅𝑝 and (𝐴, 𝐶) observable.

– Let 𝜓 the known nonlinear output feedback Lipschitz.
• The nonlinear observer is designed as:

– Assuming full information on 𝜓, the error dynamics 𝑥෤ = 𝑥 − 𝑥ො is:

– Design 𝐻 such that (𝐴 − 𝐻𝐶) is Hurwitz.

– Then lim 𝑥෤ 𝑡 = 0 for all initial conditions 𝑥(0).
෤
𝑡→∞

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 30
Nonlinear Observers
• Global Observers
 Consider a nonlinear Lipschitz system
• Consider a more general system:

– where, 𝜙 𝑥, 𝑢 is globally Lipschitz, uniformly in 𝑢:

• The nonlinear observer is designed as:

– Design 𝐻 such that (𝐴 − 𝐻𝐶) is Hurwitz, the error dynamics is:

– Does the estimation error 𝑥෤ → 0 as 𝑡 → ∞?

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 31
Nonlinear Observers
• Global Observers
• Consider a Lyapunov candidate 𝑉 = 𝑥෤ 𝑇 𝑃𝑥෤ :
– where, 𝑃 is the pos.def. solution to Lyapunov equation:
𝑃 𝐴 − 𝐻𝐶 + 𝐴 − 𝐻𝐶 𝑇 𝑃 = −𝐼
• Then:

– Hence the origin is globally exponentially stable if,

– Since the bound 𝐿 depends on 𝑃, which in turns depends on 𝐻,

– Design high-gain 𝐻 to make 1/(2 𝑃 ) as large as possible.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 32
Contents
Local Observers
1 Motivation, review of linear observers, nonlinear system, fixed
observer gain, optimal observer (LQE), example.

Extended Kalman Filter (EKF)

2 Time-varing observer gain, linearization, differential Riccati
equation, assumptions, theorems, example.

Global Observers
3 Nonlinear but known systems, nonlinear Lipschitz systems,
extension of linear observer design.

High-Gain Observer
4
Motivating example; 2nd order system, nonlinear observer,
rejecting uncertainty, peaking phenomenon. Observer
formulation.

In this chapter we review design of observers for nonlinear systems. First a review on linear observer designs are given, and the optimal Kalman
filter formulation is introduce. Then the use of this design for linearized system is given, and it is generalized to extended Kalman filter with
variable observer gain. Then design on nonlinear observer for known and uncertain system is elaborated.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 33
Global Observers
• High-Gain Observer
 Motivate Example:
• Consider the two-dimensional system:
𝑥ሶ 1 = 𝑥2 , 𝑥ሶ 2 = 𝜙 𝑥, 𝑢 , 𝑦 = 𝑥1
– The nonlinear dynamics 𝜙 is unknown, but its nominal value 𝜙0 is given.
• Consider the observer as:
𝑥ොሶ = 𝑥ො2 + ℎ1 𝑦 − 𝑥ො1 , 𝑥ොሶ = 𝜙0 𝑥,
ො 𝑢 + ℎ2 (𝑦 − 𝑥ො1 )
– Assume that by the choice of 𝜙0 :
𝜙0 𝑧, 𝑢 − 𝜙(𝑥, 𝑢) ≤ 𝐿 𝑥 − 𝑧 + 𝑀
– The estimation error dynamics is:

– and 𝛿 𝑥, 𝑥,
෤ 𝑢 = 𝜙 𝑥, 𝑢 − 𝜙0 (𝑥,
ො 𝑢).
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 34
High-Gain Observer
 Motivate Example:
• This is a perturbation to the linear system.
𝑇
– In the presence of 𝛿, design 𝐻 = ℎ1 , ℎ2 to rejects 𝛿 effect on 𝑥.
෤
• Find transfer matrix from 𝛿 to 𝑥.
෤

– We wish 𝐺0 (𝑠) was identically zero!

– Use high gain observer ℎ2 ≫ ℎ1 ≫ 1 to make sup 𝐺0 (𝑗𝜔) small.
𝜔∈𝑅
– Set

– for some positive constant 𝛼1 , 𝛼2 and 𝜀 with 𝜀 ≪ 1. Then

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 35
High-Gain Observer
 Motivate Example:
• This results in
lim 𝐺0 𝑠 = 0
𝜀→0
– Use singularly perturbed states: (scaling time domain)

𝑇
– Call 𝜂 = 𝜂1 , 𝜂2 the fast states.
– Then

– The matrix 𝐹 is Hurwitz, since 𝛼1 , 𝛼2 > 0.

– The eigenvalues of 𝐴0 are 1/𝜀 times slower than that of 𝐹.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 36
High-Gain Observer
 Motivate Example:
• Consider the bound 𝛿 ≤ 𝐿 𝑥෤ + 𝑀 ≤ 𝐿 𝜂 + 𝑀
• And the Lyapunov candidate 𝑉 = 𝜂𝑇 𝑃𝜂
– where 𝑃 is the solution to 𝑃𝐹 + 𝐹 𝑇 𝑃 = −𝐼,
– Then, for 𝜀𝐿 𝑃𝐵 ≤ 1/4

– For 𝜀𝐿 𝑃𝐵 ≤ 1/4:

– Hence, 𝑥෤ is UUB by 𝜀𝑐𝑀 for some 𝐶 > 0, and

– For some positive constant 𝑎, 𝑘.

– The smaller 𝜀 the faster the rate of decay 𝜂 → UUB.
– If 𝑀 = 0 (𝜙 = 𝜙0 ), then 𝑥෤ → 0.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 37
High-Gain Observer
 Motivate Example:
• Peaking Phenomenon:
– Since 𝜂1 = 𝑥/𝜀෤ if 𝑥1 0 ≠ 𝑥ො1 (0) then 𝜂1 0 = 𝑂(1/𝜀)!
– The solution to 𝜂 has a term of (1/𝜀)𝑒 −𝑎𝑡/𝜀
– A fast transient with a peaking amplitude: (1/𝜀) → ∞ as 𝜀 → 0.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 38
High-Gain Observer
• Observer Formulation:
 Consider the nonlinear system on the form:

– where 𝑥 ∈ 𝑅𝜌 , 𝑢 ∈ 𝑅𝑚 , 𝑦 ∈ 𝑅 are the state, input and output, respectively.

– The disturbance 𝜔 ∈ 𝑅𝜌 may have a known dynamics.
– The term 𝜓𝑖 is the known nonlinear dynamics, and locally Lipschitz:

– The term 𝜙 is the uncertain nonlinear dynamics, with nominal value 𝜙0 .

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 39
High-Gain Observer
• Observer Formulation:
 Assume bounded 𝑥 𝑡 , 𝑢 𝑡 , 𝜔 𝑡 .
• The high-gain observer is defined by:

– where 𝜀 > 0 is sufficiently small constant.

– 𝛼1 to 𝛼𝑝 are chosen to have Hurwitz polynomial.

– Note that the observer gains 𝛼𝑖 /𝜀 𝑖 have order of magnitude

increasing gain by increasing 𝑖.
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 40
High-Gain Observer
• Observer Formulation:

• Proof: Define the scaled estimation errors:

• where 𝜂 satisfies:
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 41
High-Gain Observer
• Observer Formulation:
• Proof: (Cont.)
T
– In which 𝛿 = 𝛿1 , 𝛿2 , … , 𝛿𝜌 ,

– and

– The matrix is Hurwitz by design, and (11.32)/

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 42
High-Gain Observer
• Observer Formulation:
• Proof: (Cont.)
– Use (11.29)

– This and (11.33) shows that

– Independent of 𝜀 > 0.
– Let 𝑉 = 𝜂 𝑇 𝑃𝜂, with 𝑃 being the solution of 𝑃𝐹 + 𝐹 𝑇 𝑃 = −𝐼, then

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 43
High-Gain Observer
• Observer Formulation:
• Proof: (Cont.)
– Use (11.37) to find:

– For

– Therefore, by UUB stability Theorem:

– From (11.35) we see that and

– This yields to

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 Nonlinear Control
Chapter 7: Nonlinear Observers
To read more and see the course videos
visit our course website:
http://aras.kntu.ac.ir/arascourses/nonlinear-control/

Thank You
Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,
Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020
 About Hamid D. Taghirad
Hamid D. Taghirad has received his B.Sc. degree in mechanical engineering
from Sharif University of Technology, Tehran, Iran, in 1989, his M.Sc. in mechanical
engineering in 1993, and his Ph.D. in electrical engineering in 1997, both
from McGill University, Montreal, Canada. He is currently the University Vice-
Chancellor for Global strategies and International Affairs, Professor and the Director
of the Advanced Robotics and Automated System (ARAS), Department of Systems
and Control, Faculty of Electrical Engineering, K. N. Toosi University of Technology,
Hamid D. Taghirad Tehran, Iran. He is a senior member of IEEE, and Editorial board of International
Journal of Robotics: Theory and Application, and International Journal of Advanced
Professor Robotic Systems. His research interest is robust and nonlinear control applied to
robotic systems. His publications include five books, and more than 250 papers in
international Journals and conference proceedings.

Nonlinear Control K. N. Toosi University of Technology, Faculty of Electrical Engineering,

Prof. Hamid D. Taghirad Department of Systems and Control, Advanced Robotics and Automated Systems October 21, 2020