Outline Sliding Control Continuous Approximations of Switching Control Laws

Nonlinear Control
Lecture 11: Sliding Control

Farzaneh Abdollahi

Department of Electrical Engineering

Amirkabir University of Technology

Fall 2011

Farzaneh Abdollahi Nonlinear Control Lecture 11 1/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Sliding Control
Sliding Surface
Integral Control
Gain Margins

Continuous Approximations of Switching Control Laws

Farzaneh Abdollahi Nonlinear Control Lecture 11 2/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Sliding Control
I Sliding control is a robust control technique to control systems with
model imprecision and uncertainties.
I Sliding control is based on the idea that ”controlling a 1st order system is
much easier than the general nth order system
I To achieve this goal:
1. A first order system (sliding surface) is proposed and provide a condition
(sliding condition) to make the introduced surface an invariant set of the
system stability
2. A control is designed to reach to the sliding surface
I Providing perfect performance in presence of arbitrary parameter
inaccuracy is at the price of extremely high control activity.
I ∴ a modification of control law is required to provide an effective
trade-off between tracking performance and parametric uncertainty.
I In some specific applications, such as those involving the control of
electric motor the unmodified control law can be applied directly.
Farzaneh Abdollahi Nonlinear Control Lecture 11 3/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

Sliding Surface
I Consider single input dynamics
x (n) = f (x) + b(x)u (1)
I f is not exactly known, upper bounded by known continuous function of x
I b is not exactly known, its sign is known and upper bounded by known
continuous function of x
(n−1) T
I Objective: find u, s.t. x track xd = [xd , ẋd , . . . , xd ] in presence of
imprecision on f (x) and b(x)
I Tracking error vector: x̃ = x − xd = [x̃ x̃˙ . . . x̃ (n−1) ]
I Define a time-varying surface S(t) in state-space R n by scaler equation
s(x; t) = 0: d
s(x; t) = ( + λ)n−1 x̃ (2)
dt
where λ > 0 conts.
I for n = 2 s = x̃˙ + λx̃, s is a weighted sum of position error and velocity
error
I for n = 3 s = x̃¨ + 2λx̃˙ + λ2 x̃
Farzaneh Abdollahi Nonlinear Control Lecture 11 4/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

Sliding Surface

I The problem of tracking the n-dimensional vector xd (the original

tracking problem) can be replaced by a 1st-order stabilization
problem in s.
I Given initial condition xd (0) = x(0), the problem of tracking x ≡ xd is
equivalent to remaining on the surface S(t) for all t > 0 (s ≡ 0 represents
a linear differential equation whose unique solution is x̃ ≡ 0)
I In (1),s contains x̃(n − 1) we only need to differentiate s once for the
input u to appear.
I Bounds on s can be directly translated into bounds on x̃ s represents a
true measure of tracking performance. When x̃(0) = 0:

∀t ≥ 0, |s(t)| ≤ Φ ⇒ ∀t ≥ 0, |x̃ (i) | ≤ (2λ)i ε, i = 0, ..., n − 1 (3)

where ε = Φ/λn−1

Farzaneh Abdollahi Nonlinear Control Lecture 11 5/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

I Proof: x̃ is obtained from s through a sequence of first-order lowpass filters,

shown in Fig.

Rt
I Let y1 output of first filter: y1 = 0 e −λ(t−T ) s(T )dT , |s| ≤ Φ ⇒
R t −λ(t−T )
|y1 | ≤ Φ 0 e dT = (Φ/λ)(1 − e −λt ) ≤ Φ/λ
I Repeat the same procedure all the way to yn−1 = x̃ |x̃| ≤ Φ/λn−1 = ε
I To obtain x̃ (i) , see the Fig b

I The output of the (n − 1 − i)th filter: z1 < Φ/λn−1−i

p±λ λ λ
I Note that p+λ =1− λ+p ≤1+ λ+p
I ∴|x̃ (i) | ≤ (Φ/λn−1−i )(1 + λλ )i = (2λ)i ε
I If x̃(0) 6= 0, , (3) is obtained asymptotically, within a short time-constant
(n − 1)/λ.
Farzaneh Abdollahi Nonlinear Control Lecture 11 6/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

Sliding Condition
I To keep the scalar s at zero, a control law u should
be found s.t outside of S(t):

1 d 2
s ≤ −η|s| (4)
2 dt
where η > 0 conts.
I ∴ The squared ”distance” to the surface, s 2 ,
decreases along all system trajectories. (V = 12 s 2 )
I (4), so-called sliding condition, makes the surface an
invariant set.
I By keeping the invariant set, some disturbances or
dynamic uncertainties can be tolerated.
I S(t) is sliding surface; behavior of the system on the
surface is sliding mode

Farzaneh Abdollahi Nonlinear Control Lecture 11 7/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

I If it is on the sliding surface, the system behavior can be expressed by

d
( dt + λ)n−1 x̃ = 0
I If sliding condition is guaranteed, for nonzero initial condition,
(x(0) 6= xd (0)), the surface S(t) will be reached in a finite time smaller
than |s(t = 0)|/η:
I For treach : required time to reach s = 0, integrate (4) from 0 to treach :
s(treach ) − s(0) = 0 − s(0) < −η(treach − 0) treach ≤ |s(t = 0)|/η
I Once on the surface, tracking error tends exponentially to zero with time
constant (n − 1)/λ
I from the sequence of (n − 1) filters of time constants equal to 1/λ

Farzaneh Abdollahi Nonlinear Control Lecture 11 8/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Local Asymptotic Stabilization

I For n = 2
I sliding surface is a line with slope −λ
I Starting with any initial conditions, the traj. reaches the time-varying
surface in finite time ≤ |s(t = 0)|/η
I Then slide along the surface towards xd exp. with time constant 1/λ

Farzaneh Abdollahi Nonlinear Control Lecture 11 9/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

I After defining the sliding surface s, the control is designed in two steps
1. A feedback control law u is selected so as to verify sliding condition (4)
2. The discontinuous control law u is suitably smoothed to achieve an optimal
trade-off between control bandwidth and tracking precision
I To cope with modeling imprecision and disturbances, the control law has to
be discontinuous across S(t).
I Implementing the associated control switchings is always imperfect
(switching is not instantaneous, and the value of s is not known with
infinite precision) yields chattering
I Chattering high control activity and may excite high frequency dynamics
neglected in modeling (such as unmodeled structural modes, neglected
time-delays, and so on).

Farzaneh Abdollahi Nonlinear Control Lecture 11 10/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Example
I Consider ẍ = f + u (5)

I f is unknown,but estimated by fˆ, estimation error on f assumed to be bounded

by known function F = F (x, ṫ) |fˆ − f | ≤ F
I To track x ≡ xd , define the sliding surface:
d
s = ( + λ)x̃ = x̃˙ + λx̃ ṡ = f + u − ẍd + λx̃˙
dt
I Best approximation û to achieve ṡ = 0
û = −fˆ + ẍd − λx̃˙

I The feedback control strategy is chosen intuitive ”if the error is negative, push
hard enough in the positive direction (and conversely)”
I To satisfy (4), a term discontinuous across the surface s = 0:
u = û − ksgn(s)

sgn(s) = 1 if s > 0
where
sgn(s) = −1 if s < 0
Farzaneh Abdollahi Nonlinear Control Lecture 11 11/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

I Note that this strategy works only for first-order systems.

I By choosing k to be large enough (4) can be guaranteed

1d 2
s = ṡ.s = (f − fˆ)s − k|s|
2 dt
’
I letting k = F + η 12 dt
d 2
s ≤ −η|s|
I Integral Control: To minimize the R t reaching time and make s(t = 0) = 0,
one can use integral control, i.e. 0 x̃(r )dr as variable of interest.
I The previous example
Rt is third order relative
R tto this variable, so s:
d
s = ( dt + λ)2 ( 0 x̃(r )dr ) = x̃˙ + 2λx̃ + λ2 0 x̃(r )dr
I The approximation of control law will be changed to
û = −fˆ + ẍd − 2λx̃˙ − λ2 x̃

I The control law, u and k will remainRthe same

t
I Now if x̃(0) 6= 0 s = x̃˙ + 2λx̃ + λ2 0 x̃(r )dr − x̃(0)
˙ − 2λx̃(0)
I ∴ Although x̃(0) 6= 0, s(t = 0) = 0
Farzaneh Abdollahi Nonlinear Control Lecture 11 12/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

Gain Margins
I Consider ẍ = f + bu

where the control gain , b which is may be time-varying or

state-dependent is unknown, but of known bounds
0 < bmin ≤ b ≤ bmax

I choose estimation of b as its geometric mean of bounds: b̂ = (bmin bmax )1/2 .

I ∴ β −1 ≤ bb̂ ≤ β,
I β = (bmax /bmin )1/2 is gain margin
I With s and û defined in previous example u = b̂ −1 [û − ksgn(s)]
I ṡ = (f − b b̂ −1 fˆ) + (1 − b b̂ −1 )(−ẍd + λx̃)
˙ − b b̂ −1 ksgn(s)
I ∴ to satisfy sliding condition
k ≥ |b̂b −1 f − fˆ + (b̂b −1 − 1)(−ẍd + λx̃)|
˙ + η b̂b −1
I Since f = fˆ + (f − fˆ), where |f − fˆ| ≤ F k ≥ β(F + η) + (β − 1)|û|
Farzaneh Abdollahi Nonlinear Control Lecture 11 13/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

Continuous Approximations of Switching Control Laws

I For system dynamics (1) a unique smooth
control to track a feasible trajectory is
u(t) = b(xd )−1 [ẍd − f (xd )]
I Control laws obtained by using sliding control
which provides ”perfect” tracking in the face
of model uncertainty, are discontinuous across
the surface S(t), chattering.
I In general, chattering is undesirable, since it
causes high control activity, and may excite
high-frequency dynamics neglected in
modeling
I The chattering is avoided by smoothing out
the control discontinuity in a thin boundary
layer neighboring the switching surface
B(t) = {X , |s(x; t)| ≤ Φ}, Φ > 0 is the
boundary layer thickness
ε = Φ/λn−1 is the boundary width
Farzaneh Abdollahi Nonlinear Control Lecture 11 14/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

I Outside of B(t), the control law u is like

before to guarantee that the boundary
layer is invariant
I All trajectories starting inside B(t = 0)
remain inside B(t) for all t > 0
I Inside B(t), u is interpolated
I For instance,, inside B(t), in the
expression of u replace sgn(s) by s/Φ, as
shown in Fig
I As it has been shown before, instead of
perfect tracking, tracking to within a
guaranteed precision εis guaranteed.
I For all trajectories starting inside
B(t = 0)
∀t ≥ 0|x̃ (i) | ≤ (2λ)i ε i = 1, ..., n − 1

Farzaneh Abdollahi Nonlinear Control Lecture 11 15/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Example

I Consider the system dynamics

ẍ + a(t)ẋ 2 cos 3x = u
I 1 ≤ a(t) ≤ 2, for simulation a(t) = | sin t| + 1,
I λ = 20, η = 0.1
I fˆ = 1.5ẋ 2 cos 3x, F = 0.5ẋ 2 | cos 3x|
I By using the switching control law: u = û − ksgn(s) =
1.5ẋ 2 cos3x + ẍd − 20x̃˙ − (0.5ẋ 2 | cos 3x| + 0.1)sgn(x̃˙ + 20x̃)

Farzaneh Abdollahi Nonlinear Control Lecture 11 16/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Example Cont’d

I Tracking performance is excellent at the price of high control chattering

Farzaneh Abdollahi Nonlinear Control Lecture 11 17/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Example Cont’d
I Modify control law by considering a thin boundary layer of thickness 0.1
I u = û − ksat(s/Φ) =
1.5ẋ 2 cos3x + ẍd − 20x̃˙ − (0.5ẋ 2 | cos 3x| + 0.1)sat((x̃˙ + 20x̃)/0.1)

I The tracking is not as perfect as before but acceptable, instead the

control law is smooth
Farzaneh Abdollahi Nonlinear Control Lecture 11 18/27
 Outline Sliding Control Continuous Approximations of Switching Control Laws

I The smoothing of control discontinuity inside B(t) actually assigns a low

pass filter structure to the local dynamics of the variables to eliminating
chattering
I Recognizing this filter-like structure allows us to
I tune up the control law by selecting λ and Φ properly s.t achieve a trade-off
between tracking precision and robustness to unmodeled dynamics.
I Φ can be made time varying
I Case 1: b = b̂ = 1
I Φ is TV the sliding condition (4) to guarantee the decreasing distance
to the boundary layer is changed to:
1 d 2
ksk ≥ Φ : s ≤ (Φ̇ − η)|s| (6)
2 dt

I The boundary layer attraction ↑ when the boundary layer ↓ (Φ̇ < 0)
I The boundary layer attraction ↓ when the boundary layer ↑ (Φ̇ > 0)

Farzaneh Abdollahi Nonlinear Control Lecture 11 19/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Case 1: b = b̂ = 1

I the control signal is modified as:

u = û − k̄(x)sat(s/Φ)
 − Φ̇
k̄(x) = k(x)
I

y if |y | ≤ 1
I sat(y ) =
sgn(y ) otherwise
I So the system trajectories inside the boundary layer:
ṡ = −k̄(x) Φs − ∆f (x) = −k̄(xd ) Φs + (−∆f (xd ) + O(ε))
where ∆f = fˆ − f
I We can consider a first order filter:
I its dynamic depends on desired state xd
I s : a measure of the algebraic distance to the surface S(t) is its output
I the ”perturbations,” (uncertainty ∆f (xd )) is its input

Farzaneh Abdollahi Nonlinear Control Lecture 11 20/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

I s provides tracking error x̃ by further low pass filtering (2)

I λ is break-frequency of the filter
I It must be chosen to be ”small” with respect to high-frequency unmodeled
dynamics (such as unmodeled structural modes or neglected time delays)
k̄(xd )
I Let us define Φ based on bandwidth λ: Φ =λ
I and:

φ̇ + λΦ = k(xd ) (7)
k̄(x) = k(x) − k(xd ) + λΦ

Farzaneh Abdollahi Nonlinear Control Lecture 11 21/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

I The boundary layer thickness Φ is defined based on the evolution of

dynamic model uncertainty
I Control signal depends on s
I s-trajectory represents a TV measure of the validity of the assumptions
on model uncertainty
I tracking error x̃ is a filtered version of s

Farzaneh Abdollahi Nonlinear Control Lecture 11 22/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Example

I Recall the previous example and modify the control properly:

I u = 1.5ẋ 2 cos3x + ẍd − 20x̃˙ − (0.5ẋ 2 | cos 3x| + η+Φ̇)sat((x̃˙ + 20x̃)/Φ)
Φ̇ = −λΦ + 0.5x˙d 2 | cos 3xd | + η
η
I ẋd (0) = 0, η = 0.1, λ = 20, Φ(0) = λ
I Max of the Φ is the same as the constant value of Φ in previous example
I The tracking error is about 4 times better

Farzaneh Abdollahi Nonlinear Control Lecture 11 23/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Example Cont’d

Farzaneh Abdollahi Nonlinear Control Lecture 11 24/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Case 2: β 6= 1

b(xd )
I Define: βd = β(xd ) = b̂(xd )
λΦ
I If k(xd ) ≥ βd ⇒Φ̇ + λΦ = βd k(xd )
λΦ λΦ k(xd )
I If k(xd ) ≤ βd ⇒Φ̇ + βd2
= βd
I Φ(0) = βd k(xd (0))/λ
k̄(xd )βd
I Modify λ = Φ
λΦ
I And finally k̄(x) = k(x) − k(xd ) + βd

Farzaneh Abdollahi Nonlinear Control Lecture 11 25/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Remarks

1. The desired trajectory xd must be smooth enough not to excite the

high-frequency unmodeled dynamics.
2. The sliding control guarantees the best tracking performance given the
desired control bandwidth and the extent of parameter uncertainty.
3. If the model or its bounds are so imprecise that F can only be chosen as
a large constant, then define Φ a large constant, s.t. the term
k̄sat(s/Φ) = λs/β like simple P.D.

Farzaneh Abdollahi Nonlinear Control Lecture 11 26/27

 Outline Sliding Control Continuous Approximations of Switching Control Laws

Remarks

4. For exceptional disturbances which their intensity is high s.t may take
the traj. out of the boundary:
I If integral control is applied, the integral term in the control may become
unreasonably large
I once the disturbance stops, the system goes through large amplitude
oscillations in order to return to the desired trajectory (integrator windup)
I It is a potential cause of instability because of saturation effects and
physical limits on the motion.
I Solution: As long as the system is outside the boundary layer maintain
the integral term constant
I When the system remains in the boundary layer (returns to normal case
after the exceptional disturbance) integration can resume

Farzaneh Abdollahi Nonlinear Control Lecture 11 27/27