Professional Documents
Culture Documents
Adaptive Control
Basics and Research
Gang Tao
& %
1
' $
w(t)
Feedback
& %
2
' $
• System modeling
• Control objectives
stability, transient, tracking, optimality, robustness
• Parametric uncertainties
payload variation, component aging, condition change
• Structural uncertainties
component failure, unmodeled dynamics
• Environmental uncertainties
external disturbances
• Nonlinearities
smooth functions and nonsmooth (“hard”) characteristics
& %
3
' $
• Real-time implementable
• State variables
x, y, z = position coordinates φ = roll angle
u, v, w = velocity coordinates θ = pitch angle
p = roll rate ψ = yaw angle
q = pitch rate β = side-slip angle
r = yaw rate α = angle of attack
& %
5
' $
– Force equations:
L, M, N: aerodynamic torques
& %
6
' $
& %
7
' $
ym (t)
- Reference model
θ(t) ?−
Adaptive law
θ̇(t)
6
6 6
?
r(t) u(t) y(t)
- Controller - -
Plant
C(s; θ(t))
6
& %
8
' $
θ p (t)
Design Parameter estimator
equation θ̇ p (t)
θc (t) 6 6
?
ym (t) u(t) y(t)
- Controller - Plant -
C(s; θc (t)) G(s; θ∗p )
6
& %
9
' $
• G. Tao, Adaptive Control Design and Analysis, John Wiley & Sons,
Hoboken, New Jersey, 2003.
& %
10
' $
& %
11
' $
& %
12
' $
& %
13
' $
& %
14
' $
& %
' $
& %
' $
& %
1
' $
& %
2
' $
Summary
• System uncertainties
– common in control systems
– challenges for system performance
• Adaptive control
– handles system uncertainties effectively
– ensures desired asymptotic performance
• Adaptive control theory
– mature with systematic design procedures
– developing with new challenges
• Adaptive control techniques
– proved to be useful for many practical control problems
– promising for new aerospace applications
& %
3
' $
& %
4
' $
• Research motivation
& %
5
' $
• Dead-zones
hydraulic valves, servo motors
• Backlash
gear-boxes, mechanical links
• Hysteresis
piezoelectric materials
• Piecewise-linearity
different operation conditions
• Non-linear characteristics
non-uniform hardware properties
& %
6
' $
Return
Piston
Pressure
source
Return
M, f y(t)
Load
Figure 1: Dead-zone in a servo-valve.
& %
7
' $
6
v u y
- - - K -
s(Ms+B)
& %
8
' $
v(t)
u(t)
-cl cr
& %
9
' $
m
cl
0 cr v
v control
gear-train
backlash
u
h(s) = G(s)(u(s) - d(s))
u = B(v) G(s) = k/s
& %
11
' $
ym Amplifier y
Controller Gear
and train
- motor
Mirror
& %
12
' $
ωl ωm
bl bm
J Jm u
l
θl θm
(a)
-b b -b b
(b) (c)
Figure 7: (a) Gear-train with backlash and flexibility; (b) Backlash for rigid gears:
θl = B(θm ); (c) Dead-zone: δ = D(θm − θl ).
& %
13
' $
f- f- 1 ω- θ- δ ω-l θl
−
u ? v - f-
- m 1 m f-
D(·) - k
1 1 -
− 6 Jm s s −6 −6 Jl s s
bm bl
& %
14
' $
Research Motivation
& %
16
' $
r ud v u y
- i - C1 (D) - NI
a (·) - Na (·) - G(D) -
−
6
C2 (D)
An Illustrative Example
r v u y
- k - α + βs - B(·) - 1
s
-
−6
PI controller Plant
& %
19
' $
4 6
2
4
0
2
-2
0
-4
-6 -2
-2 0 2 4 -2 0 2 4
(a) e(t) vs. v(t). (b) u(t) vs. v(t).
-2
-4
-6
0 10 20 30 40 50
(c) Tracking error e(t).
4
0
3
−2 2
1
−4
0
−6 −1
−2 0 2 4 −2 0 2 4
(a) e(t) vs. v(t). (b) u(t) vs. v(t).
2 1
0.5
−2
−4
0
−6
0 10 20 30 40 0 10 20 30 40
(c) Tracking error e(t). ^
(d) Parameter error c(t) − c.
y- zr u- v u z y
r
NIo (·) - g d
NIi (·) - Ni (·) - -
G(s) No (·) -
CO
6
C
C θ∗T a (s)
1 λ
z̄
θ∗T
2 bλ (s)
NIo (·)
& %
22
' $
r ud w v u z y
- θ23 - g - NI(·) - - θ13 - g - Gi (s) - N(·) - Go (s) -
66
CO 6
CO
C C
C θT ω21 C θT ω11
21 11
θT22 ω22
& %
23
' $
r ud w v u z y
- θ23 - g - - NI(·) - θ13 - g - Gi (s) - N(·) - Go (s) -
66
CO 6
CO
C C
C θT ω21 C θT ω11
21 11
θT22 ω22
& %
24
' $
• Research motivation
& %
25
' $
Research Motivation
& %
26
' $
& %
28
' $
• Control objective
adaptive compensation of N(·; θ∗ )
tracking control for aircraft flight trajectory
• Design strategy
adaptive inversion of N(·; θ∗ )
feedback control for aircraft dynamics
& %
29
' $
Variation of deflection with actuation voltage (θ∗2 = 15)
15
*
θ1 = 20
10 *
θ = 30
1
u(t), the virtual deflection on the airfoil
*
θ = 40
5 1
−5
−10
−15
−20
−25
0 2 4 6 8 10 12
v(t), the input voltage in volts
& %
30
' $
& %
31
' $
• Nonlinearity parametrization
∗ ∗T 1
u(t) = N(θ ; v(t)) = −θ ω(t), ω(t) = [ , −1]T
v(t)
• Nonlinearity inverse
d (t)) = θ1 (t)
v(t) = NI(u
θ2 (t) − ud (t)
• Control error
θ2 (t) − ud (t)
u(t) − ud (t) = (θ1 (t) − θ∗1 ) − (θ2 (t) − θ∗2 )
θ1 (t)
= (θ(t) − θ∗ )T ω(t)
& %
32
' $
• System model
ẋ(t) = Ax(t) + Bu(t)
• Control error
u(t) − ud (t) = (θ(t) − θ∗ )T ω(t)
• Reference model
ẋr (t) = (A − BK)xr (t) + Br(t)
• Error system
Simulation Results
• Control gain K
– LQRdesign with Q = I4 , R = 10
–K= 1.0113 −77.1793 115.8959 −9.1691
0.751 14.980 −159.812 8.2617
14.980 −138979.668
27181.878 7843.345
P=
−159.813 −138979.668 723352.800 −40670.052
8.262 7843.345 −40670.052 2301.187
• Reference signal:
1.5 sin(t) 0 ≤ t ≤ 60
r(t) =
1.5 sin(t) + 3 sin(2t) t ≥ 60
• Adaptation gains: γ1 = 1, γ2 = 2
& %
36
' $
Simulation I: Adaptive inverse performance
10
ft/sec
0
−10
0 50 100 150 200
Plant state x (t) (solid), reference state x (t) (dotted) vs. time (sec)
1 m1
0.1
deg/sec
0
−0.1
0 50 100 150 200
Plant state x2(t) (solid), reference state xm2(t) (dotted) vs. time (sec)
0.05
deg/sec
0
−0.05
0 50 100 150 200
Plant state x3(t) (solid), reference state xm3(t) (dotted) vs. time (sec)
0.5
deg
−0.5
0 50 100 150 200
Plant state x4(t) (solid), reference state xm4(t) (dotted) vs. time (sec)
10
ft/sec
0
−10
0 50 100 150 200
Tracking error e1(t) vs. time (sec)
0.2
deg/sec
0
−0.2
0 50 100 150 200
Tracking error e2(t) vs. time (sec)
0.05
deg/sec
0
−0.05
0 50 100 150 200
Tracking error e3(t) vs. time (sec)
0.5
deg
−0.5
0 50 100 150 200
Tracking error e4(t) vs. time (sec)
ft/sec
−5
−10
0 50 100 150 200
Tracking error e (t) vs. time (sec)
1
0.1
deg/sec
0
−0.1
0 50 100 150 200
Tracking error e2(t) vs. time (sec)
0.1
deg/sec
0
−0.1
0 50 100 150 200
Tracking error e3(t) vs. time (sec)
0
deg
−0.5
−1
0 50 100 150 200
Tracking error e4(t) vs. time (sec)
• Simulation signals
(i) r1 (t) = 5r(t) (convergent)
(ii) r2 (t) = 6r(t) (non-covergent)
both leading to saturation of ud (t).
& %
40
' $
(i) convergent results
50
ft/sec
0
−50
0 50 100 150 200
Plant state x (t) (solid), reference state x (t) (dotted) vs. time (sec)
1 m1
0.5
deg/sec
0
−0.5
0 50 100 150 200
Plant state x2(t) (solid), reference state xm2(t) (dotted) vs. time (sec)
0.1
deg/sec
0
−0.1
0 50 100 150 200
Plant state x3(t) (solid), reference state xm3(t) (dotted) vs. time (sec)
0.5
deg
−0.5
0 50 100 150 200
Plant state x4(t) (solid), reference state xm4(t) (dotted) vs. time (sec)
10
ft/sec
0
−10
0 50 100 150 200
Tracking error e1(t) vs. time (sec)
0.2
deg/sec
0
−0.2
0 50 100 150 200
Tracking error e2(t) vs. time (sec)
0.05
deg/sec
0
−0.05
0 50 100 150 200
Tracking error e (t) vs. time (sec)
3
0.5
deg
−0.5
0 50 100 150 200
Tracking error e4(t) vs. time (sec)
50
ft/sec
0
−50
0 50 100 150 200
Plant state x (t) (solid), reference state x (t) (dotted) vs. time (sec)
1 m1
1
deg/sec
0
−1
0 50 100 150 200
Plant state x2(t) (solid), reference state xm2(t) (dotted) vs. time (sec)
0.2
deg/sec
0
−0.2
0 50 100 150 200
Plant state x3(t) (solid), reference state xm3(t) (dotted) vs. time (sec)
0.5
deg
−0.5
0 50 100 150 200
Plant state x4(t) (solid), reference state xm4(t) (dotted) vs. time (sec)
20
ft/sec
0
−20
0 50 100 150 200
Tracking error e1(t) vs. time (sec)
0.5
deg/sec
0
−0.5
0 50 100 150 200
Tracking error e2(t) vs. time (sec)
0.1
deg/sec
0
−0.1
0 50 100 150 200
Tracking error e (t) vs. time (sec)
3
0.5
deg
−0.5
0 50 100 150 200
Tracking error e4(t) vs. time (sec)
Discussion
& %
45
' $
• Research motivation
& %
46
' $
Research Motivations
• Actuator failures
– common in control systems
– uncertain in failure time, pattern, parameters
– undesirable for system performance
• Adaptive control
– deals with system uncertainties
– ensures desired asymptotic performance
– is promising for actuator failure compensation
– has potential for critical applications
& %
47
' $
• Effective methods for handling system failures
– multiple-model, switching and tuning
– indirect adaptive control
– fault detection and diagnosis
– robust or neural control
• Direct adaptive failure compensation approach
– use of a single controller structure
– direct adaptation of controller parameters
– no explicit failure (fault) detection
– stability and asymptotic tracking
• Potential applications include
– aircraft flight control
– smart structure vibration control
– space robot control
& %
48
' $
System Models
m
ẋ = f (x) + ∑ g j (x)u j , y = h(x)
j=1
m
ẋ = Ax + ∑ b j u j , y = Cx
j=1
& %
49
' $
Actuator Failures
• Loss of effectiveness
• Lock-in-place
u j (t) = ū j , t ≥ t j , j ∈ {1, 2, . . . , m}
• Lost control
• Failure uncertainties
the failure values k j , ū j and d¯jk , failure time t j , pattern j, and
components δ j (t) are all unknown.
How much, how many, which and when the failures happen??
& %
50
' $
• Examples
– aircraft aileron, stabilizer, rudder or elevator failures
their segments stuck in unknown positions
their unknown broken pieces (including wings)
– satellite motion control actuator failures
– MEM actuator/sensor failures on fairing surface
– heating device failures in material growth
– generator failures in power systems
– transmission line failures in power system
– power distribution network failures
– cooperating manipulator failures
– bioagent distribution system failures
– etc.
& %
51
' $
σ = diag{σ1 , σ2 , . . . , σm }
1 if the jth actuator failed, i.e., u j (t) = ū j (t)
σj =
0 otherwise
& %
52
' $
Block Diagram
ū1 - σ
1
- 1−σ1 -?
r- v1 .. u1 -
. .. y
.. .. ū m - σm . -
.
-
Controller . - System
vm - 1−σm -?
um
& %
53
' $
Research Goals
& %
54
' $
Technical Challenges
• Redundancy
– necessary for failure compensation
– problematic for control: up to m − q failures
• Failure uncertainties
parametric, structural, and environmental
• Application issues
– system modeling and control implementation
– aircraft, robot, smart structure, power system, satellite
& %
55
' $
Control Objective
Stability and asymptotic tracking for up to m−q failures.
Basic Assumption
The system is so constructed that for any up to m − q (0 < q ≤ m) failed
actuators, the alive actuators can still achieve some desired performance.
Key Task
Adaptively adjust the remaining actuators (controls) to achieve desired
performance when system and failure parameters are unknown.
& %
56
' $
Design Tools
& %
57
' $
& %
58
' $
• Model-based approach
for asymptotic tracking despite uncertain failures
& %
59
' $
• System model
0.08
y(t)
y(t), ym(t) (rad) 0.06 y (t)
m
0.04
0.02
0
0 10 20 30 40 50 60 70 80 90 100
time (sec)
0.015
e(t) (rad)
0.01
0.005
0
0 10 20 30 40 50 60 70 80 90 100
time (sec)
0
−1
v(t) (deg)
−2
−3
−4
0 10 20 30 40 50 60 70 80 90 100
time (sec)
& %
61
' $
0.08
y(t)
0.04
0.02
0
0 10 20 30 40 50 60 70 80 90 100
time (sec)
0.02
e(t) (rad)
0.01
−0.01
0 10 20 30 40 50 60 70 80 90 100
time (sec)
0
−1
v(t) (deg)
−2
−3
−4
0 10 20 30 40 50 60 70 80 90 100
time (sec)
& %
62
' $
ẋ = Ax + Bu, y = Cx
• Actuator failures
dr2 fails at t = 50, da2 fails at t = 100 seconds
• Simulation results
& %
63
' $
4
deg
0
0 20 40 60 80 100 120 140 160 180 200
Roll angle φ (t): −, reference output φm (t): −−
10
6
deg
0
0 20 40 60 80 100 120 140 160 180 200
Yaw angle ψ(t): −, reference output ψm (t): −−
& %
64
' $
Discussion
Conclusions
Questions?
Gang Tao
University of Virginia
& %
69