2.

153 Adaptive Control

Lecture 6
Adaptive PI Control

Anuradha Annaswamy

aanna@mit.edu

( aanna@mit.edu ) 1 / 14
Pset #1 out: Thu 19-Feb, due: Fri 27-Feb
Pset #2 out: Wed 25-Feb, due: Fri 6-Mar
Pset #3 out: Wed 4-Mar, due: Fri 13-Mar
Pset #4 out: Wed 11-Mar, due: Fri 20-Mar
Midterm (take home) out: Mon 30-Mar, due: Fri 3-Apr

( aanna@mit.edu ) 2 / 14
Adaptive Control of a Second-order Plant
+ e τ 1
Gc (s) s(Js+B)
−

Plant: J ω̈ + Bω̇ = τ J>0

ki
PI Control: Gc (s) = kp +
R s
τ = kp e(t) + ki e(τ )dτ
R
Adaptive PI Control: τ = kp (t)e(t) + ki (t) e(τ )dτ

PID Control: Gc (s) = kp + kd s + ksi

τ = kp e(t) + ki e(τ )dτ + kd de
R
dt
R
Adaptive PID Control: τ = kp (t)e(t) + ki (t) e(τ )dτ + kd (t)ė(t)
J and B are unknown. Adjust kp (t), ki (t) and kd (t) so that the
closed-loop system is stable and limt→∞ e(t) = 0.
( aanna@mit.edu ) 3 / 14
Adaptive Control of a Second-order Plant
+ e τ 1
Gc (s) s(Js+B)
−

Plant: J ω̈ + Bω̇ = τ J>0

ki
PI Control: Gc (s) = kp +
R s
τ = kp e(t) + ki e(τ )dτ
R
Adaptive PI Control: τ = kp (t)e(t) + ki (t) e(τ )dτ

PID Control: Gc (s) = kp + kd s + ksi

τ = kp e(t) + ki e(τ )dτ + kd de
R
dt
R
Adaptive PID Control: τ = kp (t)e(t) + ki (t) e(τ )dτ + kd (t)ė(t)
J and B are unknown. Adjust kp (t), ki (t) and kd (t) so that the
closed-loop system is stable and limt→∞ e(t) = 0.
( aanna@mit.edu ) 3 / 14
Adaptive Control of a Second-order Plant
+ e τ 1
Gc (s) s(Js+B)
−

Plant: J ω̈ + Bω̇ = τ J>0

ki
PI Control: Gc (s) = kp +
R s
τ = kp e(t) + ki e(τ )dτ
R
Adaptive PI Control: τ = kp (t)e(t) + ki (t) e(τ )dτ

PID Control: Gc (s) = kp + kd s + ksi

τ = kp e(t) + ki e(τ )dτ + kd de
R
dt
R
Adaptive PID Control: τ = kp (t)e(t) + ki (t) e(τ )dτ + kd (t)ė(t)
J and B are unknown. Adjust kp (t), ki (t) and kd (t) so that the
closed-loop system is stable and limt→∞ e(t) = 0.
( aanna@mit.edu ) 3 / 14
Adaptive Control of a Second-order Plant
+ e Adaptive τ 1
s(Js+B)
Controller
−

Plant: J ω̈ + Bω̇ = τ J>0

ki
PI Control: Gc (s) = kp +
R s
τ = kp e(t) + ki e(τ )dτ
R
Adaptive PI Control: τ = kp (t)e(t) + ki (t) e(τ )dτ

PID Control: Gc (s) = kp + kd s + ksi

τ = kp e(t) + ki e(τ )dτ + kd de
R
dt
R
Adaptive PID Control: τ = kp (t)e(t) + ki (t) e(τ )dτ + kd (t)ė(t)
J and B are unknown. Adjust kp (t), ki (t) and kd (t) so that the
closed-loop system is stable and limt→∞ e(t) = 0.
( aanna@mit.edu ) 3 / 14
PI -Control: Algebraic Part

r + er Nominal τ 1 x
Controller Js+B
−

ki
Gc (s) = kp + Parameterize kp = K > 0, ki = Kλ > 0
s
K(s + λ)
Closed-loop transfer function:
s(Js + B) + K(s + λ)
K(s + λ)
= 2
Js + s(B + K) + Kλ

Stable if K > |B|. Design the controller so that x → xd

( aanna@mit.edu ) 4 / 14
PI Control - Algebraic Part: Tracking
Wcl (s)

r + er τ 1 x
Gc (s) Js+B
−

Gc (s)
Wcl (s) =
Js + B + Gc (s)
Wcl−1 (s) = 1 + (Js + B)G−1
cl (s)

r = Wcl−1 (s)[xd ]
 
= xd + (Js + B)G−1
cl (s) [xd ]
= xd + (Js + B)[ωd ] = xd + Bωd + J ω̇d
( aanna@mit.edu ) 5 / 14
PI Control - Algebraic Part: Tracking
Wcl (s)

r + er τ 1 x
Gc (s) Js+B
−

Gc (s)
Wcl (s) =
Js + B + Gc (s)
Wcl−1 (s) = 1 + (Js + B)G−1
cl (s)

r = Wcl−1 (s)[xd ]
 
= xd + (Js + B)G−1
cl (s) [xd ]
= xd + (Js + B)[ωd ] = xd + Bωd + J ω̇d
( aanna@mit.edu ) 5 / 14
PI Control - Algebraic Part: Tracking
Wcl (s)

xd r + er τ x
Wcl−1 (s) Gc (s) 1
Js+B
−

Gc (s)
Wcl (s) =
Js + B + Gc (s)
Wcl−1 (s) = 1 + (Js + B)G−1
cl (s)

r = Wcl−1 (s)[xd ]
 
= xd + (Js + B)G−1
cl (s) [xd ]
= xd + (Js + B)[ωd ] = xd + Bωd + J ω̇d
( aanna@mit.edu ) 5 / 14
PI Control - Algebraic Part: Tracking
Using r = J ω̇d + Bωd + xd the block diagram can be represented as

ω̇d
J

+
ωd + r + er τ 1 x
B Gc (s) Js+B
+ −
xd

which can then be simplified to

ẋd
J

xd
B
Plant
+
xd + e + τ 1 x
Gc (s) Js+B
−

( aanna@mit.edu ) 6 / 14
PI Control - Algebraic Part: Tracking - Revised Design

J: reparametrize Gc (s): ki Kλ Ks + Kλ
Gc (s) = kp + =K+ =
s s s
(K + Jλ)s + Kλ
Change to Gc (s) =
s
(K + Jλ)s + Kλ)
Closed-loop transfer function:
s(Js + B) + (K + Jλ)s + Kλ)

Move B from feedforward - to feedback

(K + Jλ)s + Kλ
Closed-loop transfer function:
Js2 + (K + Jλ)s + Kλ

( aanna@mit.edu ) 7 / 14
PI Control - Algebraic Part: Tracking - Revised Design

ẋd
J

xd
B
Plant
+
xd + e + τ 1 x
Gc (s) Js+B
−

(K + Jλ)s + Kλ
Gc (s) =
s

( aanna@mit.edu ) 8 / 14
PI Control - Algebraic Part: Tracking - Revised Design

ẋd
J

+
xd + e + τ 1 x
Gc (s) Js+B
− +

(K + Jλ)s + Kλ
Gc (s) =
s

( aanna@mit.edu ) 8 / 14
PI Control - Algebraic Part: Tracking - Revised Design
r + er (K+Jλ)s+Kλ + τ 1 x
s Js+B
− +

(K + Jλ)s + Kλ
Wcl (s) :
Js2 + (K + Jλ)s + Kλ
r
Kλ
ωn =
J
r
K + Jλ J K + Jλ
ζ = · = √
2J Kλ 2 JKλ

Less sensitive to uncertainties in J

Less sensitive to uncertainties in B
( aanna@mit.edu ) 9 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd
J
xd + e + +τ 1 x
G c (s ) Js+B
− +
B

(K + Jλ)s + Kλ (K + Jλ)s + Kλ
Gc (s) = , Wcl (s) = 2
s Js + (K + Jλ)s + Kλ

r = Wcl−1 (s)[xd ]
 
= xd + (Js)G−1 cl (s) [xd ]

= xd + J ω̇d
τ = Bx + J ẋd + Gc (s)[e]
Z
= Bx + J ẋd + (K + Jλ)e + Kλ e(τ )dτ
 Z 
= J (ẋd + λe) + Bx + K e + λ e(τ )dτ = θ∗T φ(t)

( aanna@mit.edu ) 10 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
J
xd e2 + + τ x
1
K Js+B
+
B

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
J
xd e2 + + τ x
1
K Js+B
+
B

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
J
xd e2 + + τ x
1
K Js+B
+
B

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
J
xd e2 + + τ x
1
K Js+B
+
B

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
J
xd e2 + + τ x
1
K Js+B
+
B

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
J
xd e2 + + τ x
1
K Js+B
+
B

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
PI Control - Algebraic Part: Tracking - Complete
Design
ẋd e1
Ĵ
xd e2 + + τ x
1
K Js+B
+
B̂

τ = Bx + J ẋd + Gc (s)[e]
= Je1 (t) + Bx(t) + Ke2 (t) = θ∗T φ(t)
 Z 
e1 = (ẋd + λe) , e2 = e + λ e(τ )dτ

φ = [ e1 x e 2 ]> , θ ∗ = [ J B K ]>
Adaptive PI control:
τ = J(t)e1 + B(t)x
b b + Ke2

( aanna@mit.edu ) 11 / 14
Adaptive PI Control

τ
= b
J(t)e1 + B(t)x
b + Ke2
1
Plant+controller: ẋ = (−Bx + τ )
J
1 
= −Bx + bJ(t)e1 + B(t)x
b + Ke2
J
 Z 
e2 = e+λ e(τ )dτ

ė2 = ė + λe = ẋd − ẋ + λe
1 b 
= ẋd − Je1 + Bx
e + Ke2 + λe
J !
J
b 1 e 
= 1− e1 − Bx + Ke2
J J
K 1  
= − e2 + −e
Je1 − Bx
e − Error Model 3
J J
( aanna@mit.edu ) 12 / 14
Adaptive PI Control - Stability
ẋd e1
Ĵ
xd e2 + + τ x
1
K Js+B
+
B̂

τ= b
J(t)e1 + B(t)x
b + Ke2
K 1 e 
Error Equation: ė2 = − e2 + −Je1 − Bx
e
J J
˙e e˙ = γ1 e2 x
Adaptive Law: J = γ1 e2 e1 , B
!!
1 1 J2
e e2
B
Lyapunov function: V = e22 + +
2 J γ1 γ2
K
V̇ = − e22
J
( aanna@mit.edu ) 13 / 14
Adaptive PI Control - Stability and Asymptotic Tracking

ẋd e1
Ĵ
xd e2 + + τ x
1
K Js+B
+
B̂

e2 , e e ∈ L∞ =⇒ ė2 ∈ L∞
J, B
e2 ∈ L∞ =⇒ e ∈ L∞
e ∈ L∞ =⇒ e1 ∈ L∞
Therefore ė2 ∈ L∞ ; e2 ∈ L2
=⇒ limt→∞ e2 (t) = 0 =⇒ limt→∞ e(t) = 0

( aanna@mit.edu ) 14 / 14