Sliding Mode Observers

Professor Aamer Iqbal Bhatti

Department of Electrical Engineering

Capital University of Science and Technology, Islamabad

April 27, 2019

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 1 / 20
Linear Sliding Mode Observers - Utkin Observer

Consider a linear system described by

ẋ(t) = Ax(t) + Bu(t)

(1)
y (t) = Cx(t)
where A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n and p ≥ m . Assume that
B and C are full rank and (A, C ) is observable.
Consider the change of coordinates x 7→ Tc x whereby
 T 
Nc
Tc = (2)
C

where the columns of Nc ∈ Rn×(n−p) span the null space of C . This

transformation is nonsingular.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 2 / 20
Utkin Observer Cont..

The canonical form for the nominal system is

ẋ1 (t) = A11 x1 (t) + A12 y (t) + B1 u(t)

(3)
ẏ (t) = A21 x1 (t) + A22 y (t) + B2 u(t)
where  
x1 ln−p
Tc x = (4)
y lp
The structure of the observer is

x̂˙ 1 (t) = A11 x̂1 (t) + A12 ŷ (t) + B1 u(t) + Lν

(5)
ŷ˙ (t) = A21 x̂1 (t) + A22 ŷ (t) + B2 u(t) − ν

where (x̂1 , ŷ ) represent the state estimates, L ∈ R(n−p)×p is a gain

matrix and νi = M sgn (ŷi − yi ) where M ∈ R+

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 3 / 20
Utkin Observer Cont..

Define e1 = x̂1 − x1 and ey = ŷ − y then the error dynamics are

ė1 (t) = A11 e1 (t) + A12 ey (t) + Lν

(6)
ėy (t) = A21 e1 (t) + A22 ey (t) − ν
Since the pair (A, C ) is observable, the pair (A11 , A21 ) is also observable
and L can be chosen to make the spectrum of A11 + LA21 lie in C− .
Define a further change of coordinates by
 
ln−p L
T̃ = (7)
0 lp

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 4 / 20
Utkin Observer Cont..

With ẽ1 = e1 + Ly the error dynamics become

ẽ˙ 1 (t) = Ã11 ẽ1 (t) + Ã12 ey (t)

(8)
ėy (t) = A21 ẽ1 (t) + Ã22 ey (t) − ν

where Ã11 = A11 + LA21 , Ã12 = A12 + LA22 − Ã11 L and

Ã22 = A22 − A21 L.
In the domain
 
1 
T

Ω = (e1 , ey ) : kA21 e1 k + λmax Ã22 + Ã22 key k < M − η (9)
2

where η < M is some small positive scalar, the reachability condition

eyT ėy < −η key k is satisfied.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 5 / 20
Utkin Observer Cont..

An ideal sliding motion will take place on the surface

So = {(e1 , ey ) : ey = 0} (10)

After some finite time ts , for all subsequent time, ey = 0 and ėy = 0.
The corresponding sliding mode dynamics are given by

ẽ˙ 1 (t) = Ã11 ẽ1 (t) (11)

which, by choice of L, represents a stable system and so ẽ1 → 0 and

consequently, x̂1 → x1 as t → ∞

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 6 / 20
Tutorial Example

Consider the second-order linear system

ẋ(t) = Ax(t) + Bu(t)

(12)
y (t) = Cx(t)

where    
0 1 0  
A= B= C= 1 1
−2 0 1
This represents a simple harmonic oscillator.
For simplicity assume u = 0 and consider the problem of designing a
sliding mode observer.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 7 / 20
Tutorial Example Cont..

Define a nonsingular matrix to construct the canonical form

 
1 0
Tc = (13)
1 1

This change of coordinates gives the system triple

   
−1 1 0
Tc ATc−1 CTc−1 =
 
= Tc B = 0 1 (14)
−3 1 1

An appropriate choice of observer gain is L = 0.5 which results in an

error system governed by Ã11 = −2.5.
The scaling constant M in the discontinuous component has been set
to unity.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 8 / 20
Tutorial Example Cont..
The Figure shows the state estimation errors e1 (t) and ey (t) resulting
from the initial conditions e1 = −1 and ey = 0
Although the error system starts on the sliding surface So , an ideal slid-
ing motion cannot be maintained; only after approximately 1.2 seconds
is sliding established.
At this point in the time interval 1.2 to 3.0 seconds, e1 (t) exhibits a
first-order exponential decay to the origin.
After 3.0 seconds almost perfect replication of the states takes place.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 9 / 20
Tutorial Example Cont..

In the original coordinates perfect tracking occurs after approximately

3 seconds.
The dotted lines represent the true states and the solid line the esti-
mates from the observer.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 10 / 20
Tutorial Example Cont..

This Figure shows the value of ν with respect to time and shows switch-
ing taking place from 1.2 seconds onwards.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 11 / 20
Walcott-Zak Observer
Consider now the uncertain system described by

ẋ(t) = Ax(t) + Bu(t) + f (t, x, u)

(15)
y (t) = Cx(t)
A ∈ Rn×n , B ∈ Rn×m , C ∈ Rp×n and p ≥ m; in addition the matrices
B and C are assumed to be of full rank.
The function f : R+ × Rn × Rm → Rn is unknown and represents the
system uncertainty.
A special case is to consider a matched uncertainty
f (t, x, u) = Bξ(t, x, u) (16)
where ξ(t, x, u) is unknown, but bounded
kξ(t, x, u)k ≤ r1 kuk + α(t, y ) (17)
with r1 a known scalar and α(t, y ) a known function.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 12 / 20
Walcott-Zak Observer Cont..

The problem is to estimate the states of the uncertain system so that

the error system
e(t) = x̂(t) − x(t) (18)
is quadratically stable despite the presence of the uncertainty.
There exists a G ∈ Rn×p such that A0 = A−GC has stable eigenvalues
and there exists a Lyapunov pair (P, Q) for A0 such that the structural
constraint

C T F T = PB (19)
is satisfied for some F ∈ Rm×p

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 13 / 20
Walcott-Zak Observer Cont..

Consider an observer of the form

˙
x̂(t) = Ax̂(t) + Bu(t) − G (C x̂(t) − y (t)) + P −1 C T FνT (20)

where (
FCe
−ρ(t, y , u) kFCek if FCe 6= 0
ν= (21)
0 otherwise
and the scaler function ρ(.) is any function satisfying

ρ(t, y , u) ≥ r1 kuk + α(t, y ) + η (22)

for some positive scalar η.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 14 / 20
Walcott-Zak Observer Cont..

The error system satisfies

ė(t) = (A − GC )e(t) − Bξ(t, x, u) + Bν (23)

To prove quadratic stability of the error system, consider V (e) = e T Pe

as a candidate Lyapunov function. therefore

V̇ = e T (PA0 + A0 P) e − 2e T PBξ + 2e T PBν

(24)
≤ −e T Qe − 2e T PBξ − 2ρ(t, y , u)kFCek

Using the structural constraint C T F T = PB:

V̇ ≤ −e T Qe − 2e T C T F T ξ − 2ρ(t, y , u)kF Cek

≤ −e T Qe − 2kF Ce k(ρ(t, y , u) − kξk) (25)
≤ −e T Qe − 2ηkFCek

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 15 / 20
Walcott-Zak Observer Cont..

Hence, there exists a domain in which a sliding motion is induced on

the surface in the state error space given by

Swz = {e ∈ Rn : FCe = 0} (26)

However, it relies on establishing whether there exists a gain matrix

G such that, for the resulting closed-loop matrix A0 , there exists a
Lyapunov matrix P for A0 satisfying C T F T = PB for some F ∈ Rm×p .

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 16 / 20
Nonlinear Sliding Mode Observers - Slotine’s Observer
A Nonlinear system in companion form is written as
x (n) = f (x, t) (27)
where f (x, t) is a nonlinear, uncertain function and x1 is the measure-
ment.
The structure of the Slotine’s Observer is
x̂˙ 1 = −α1 e1 + x̂2 − k1 sgn (e1 )
x̂˙ 2 = −α2 e1 + x̂3 − k2 sgn (e1 )
.. (28)
.
ẋn = −αn e1 + fˆ − kn sgn (e1 )
αi chosen as for a Luenberger observer to ensure asymptotic error decay
when ki = 0.
e1 = x̂1 − x1 , fˆ is an estimate of f (x, t)

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 17 / 20
Slotine’s Observer Cont..

The error dynamics are

ė1 = −α1 e1 + e2 − k1 sgn (e1 )

ė2 = −α2 e1 + e3 − k2 sgn (e1 )
.. (29)
.
ėn = −αn e1 + ∆f − kn sgn (e1 )

∆f = fˆ − f is assumed bounded and kn ≥ |∆f |

The sliding condition d
dt (e1 )2 < 0 is satisfied in the region

e2 ≤ k1 + α1 e1 if e1 < 0
(30)
e2 ≥ −k1 + α1 e1 if e1 < 0

Now revisit the error dynamics imposing the sliding condition.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 18 / 20
Slotine’s Observer Cont..

The Sliding Mode Dynamics when e1 = 0

e2 − k1 sgn (e1 ) = 0 (31)

and therefore
k2
ė2 = e3 − e2
k1
.. (32)
.
kn
ėn = ∆f − e2
k1

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 19 / 20
Slotine’s Observer Cont..

The αi only affect the dynamic performance prior to the reaching of

the so-called sliding patch - patch dynamics are
  k 

 − k12 1 0 . . . 0 
 k3
− k1 0 1 . . . 0
 

λIn−1 −  .  = 0 (33)
  
  
  . . . 1
− kkn 0 0 . . . 0 
 

1

Assuming kn is selected as a constant ratio with k1 and that the poles

defining the dynamics on the patch are critically damped i.e. are real
and equal to some constant value γ, then
 
 (i) 
e2  ≤ (2γ)i k1 i = 0, . . . , n − 2 (34)

from which the precision of the state estimates can be determined.

c Dr. A. I. Bhatti, CASPR SMC Workshop April 27, 2019 20 / 20