Model Reference

Adaptive Control
Survey of Control Systems (MEM 800)

Presented by
Keith Sevcik
 Concept
ymodel
Model

Controller Parameters Adjustment

Mechanism

uc
u yplant
Controller Plant

 Design controller to drive plant response to mimic ideal

response (error = yplant-ymodel => 0)
 Designer chooses: reference model, controller structure,
and tuning gains for adjustment mechanism
 MIT Rule
 Tracking error: e  y plant  ymodel
1 2
 Form cost function: J ( )  e ( ) sensitivity
derivative
2
Update rule: d J e
    e
dt  
– Change in  is proportional to negative gradient of J
 MIT Rule
 Can chose different cost functions
EX:

J ( )  e( )
d e
  sign (e)
dt 
 1, e  0

where sign (e)   0, e  0
 1, e  0

 From cost function and MIT rule, control law can be
formed
 MIT Rule
 EX: Adaptation of feedforward gain
Reference Model

ymodel
Gm (s)  koG(s)

Adjustment Mechanism
-
θ 
Π
s +
Plant

uc u yplant
Π G p ( s)  k G( s)
 MIT Rule
Y ( s)
 For system  kG( s) where k is unknown
U ( s)
Y (s)
 Goal: Make it look like  koG ( s)
U c ( s)

using plant Gm (s)  koG(s) (note, plant model

is scalar multiplied by plant)
 MIT Rule
 Choose cost function:
1 2 d e
J ( )  e ( ) 
  e
2 dt 
 Write equation for error:
e  y  ym  kGU  GmU c  kGU c  koG U c
 Calculate sensitivity derivative:
e k
 kGUc  ym
 ko
 Apply MIT rule:
d k
   ' ym e   ym e
dt ko
 MIT Rule
 Gives block diagram:
Reference Model

ymodel
Gm (s)  koG(s)

Adjustment Mechanism
-
θ 
Π
s +
Plant

uc u yplant
Π G p ( s)  k G( s)

  considered tuning parameter

 MIT Rule
 NOTE: MIT rule does not guarantee error
convergence or stability

  usually kept small

 Tuning  crucial to adaptation rate and
stability.
 MRAC of Pendulum
 System
J  c  mgdc sin   d1  T

 ( s)
d2
d1
dc d1
 2
T ( s) Js  cs  mgdc

T
 ( s) 1.89
 2
T ( s) s  0.0389s  10.77
 MRAC of Pendulum
 Controller will take form:
ymodel
Model

Controller Parameters Adjustment

Mechanism

uc
u 1.89 yplant
Controller s  0.0389s  10.77
2
 MRAC of Pendulum
 Following process as before, write
equation for error, cost function, and
update rule:
e  y plant  ymodel
1 2
J ( )  e ( ) sensitivity
derivative
2
d J e
   e
dt  
 MRAC of Pendulum
 Assuming controller takes the form:
u  1uc   2 y plant
e  y plant  ymodel  G p u  Gmuc

1uc   2 y plant 
 1.89 
y plant  G p u   2
 s  0.0389s  10.77 
1.891
y plant  2 uc
s  0.0389s  10.77  1.89 2
 MRAC of Pendulum
1.891
e 2 uc  Gmuc
s  0.0389 s  10.77  1.89 2
e 1.89
 2 uc
1 s  0.0389 s  10.77  1.89 2
e 1.89 2 1
 uc
 2 
s  0.0389s  10.77  1.89 2
2

2

1.891
 2 y plant
s  0.0389s  10.77  1.89 2
 MRAC of Pendulum
 If reference model is close to plant, can
approximate:
s  0.0389s  10.77  1.89 2  s  a1m s  a0 m
2 2

e a1m s  a0 m
 2 uc
1 s  a1m s  a0 m
e a1m s  a0 m
 2 y plant
 2 s  a1m s  a0 m
 MRAC of Pendulum
 From MIT rule, update rules are then:
d1 e  a1m s  a0 m 
  e    2 uc e
dt 1  s  a1m s  a0 m 
d 2 e  a1m s  a0 m 
  e    2 y plant e
dt  2  s  a1m s  a0 m 
 MRAC of Pendulum
 Block Diagram
Reference Model

bm ymodel
s  a1m s  a0 m
2

-
uc + 1.89 yplant +
Π
s 2  0.0389s  10.77
-
θ1 Plant e
Π
 θ2 
s s
a1m s  a0 m a1m s  a0 m
s  a1m s  a0 m
2 Π Π s  a1m s  a0 m
2
 MRAC of Pendulum
 Simulation block diagram (NOTE: Modeled
to reflect control of DC motor)
om ega^2
s+am ym

Reference M odel

Error
Saturation
35 4.41
2/26 180/pi
s2 +.039s+10.77 y
Step Degrees
Degrees Radians
to Volts Plant to Degrees
-gam m a
s Theta1

gam m a
Theta2 s

am
am
s+am
s+am
 MRAC of Pendulum
 Simulation with small gamma = UNSTABLE!
150
ym
g=.0001

100

50

-50

-100
0 200 400 600 800 1000 1200
 MRAC of Pendulum
 Solution: Add PD feedback
om ega^2
s+am ym

Reference M odel

1.5 du/dt

D Error
Saturation
35 1 4.41
2/26 180/pi
s2 +.039s+10.77 y
Step Degrees P
Degrees Radians
to Volts Plant to Degrees
-gam m a
s Theta1

gam m a
Theta2 s

am
am
s+am
s+am
 MRAC of Pendulum
 Simulation results with varying gammas
45
ym
g=.01
g=.001
40
g=.0001

3.56 35

ym  2
s  2.67 s  3.56 30

25

Designed such that : 20

Ts  3 sec
15

  .707
10

0
0 500 1000 1500 2000 2500
LabVIEW VI Front Panel
LabVIEW VI Back Panel
Experimental Results
 Experimental Results
 PD feedback necessary to stabilize system
 Deadzone necessary to prevent updating
when plant approached model
 Often went unstable (attributed to
inherent instability in system i.e. little
damping)
 Much tuning to get acceptable response
 Conclusions
 Given controller does not perform well enough
for practical use
 More advanced controllers could be formed from
other methods
– Modified (normalized) MIT
– Lyapunov direct and indirect
– Discrete modeling using Euler operator
 Modified MRAC methods
– Fuzzy-MRAC
– Variable Structure MRAC (VS-MRAC)