2.153 adaptive control lecture 2 simple adaptive …aaclab.mit.edu/material/lect/lecture2.pdf ·...
TRANSCRIPT
2.153 Adaptive ControlLecture 2
Simple Adaptive Systems: Identification and Control
Daniel Wiese
Tuesday 17-Feb-2015
([email protected]) Tuesday 17-Feb-2015 1 / 27
Introduction
Last time:
Parameter identification
Algebraic (error model 1) and dynamic (error model 3)Scalar and vector parameter systemsNon-recursive and recursive
Stability
Introduction to using Lyapunov functions
Today:
Identification of multiple parameters in first order plant
Error model 1 and 3Determining update law using Lyapunov functions
Adaptive control
([email protected]) Tuesday 17-Feb-2015 2 / 27
Error Models
See page 273.
Relate parameter errors (cannot measure) to output/measurementerror (which can be measured)
The stability and performance of an adaptive system is dependentupon the evolution of these errors
An error model is the mathematical model which describes theevolution of these errors
By using error models, we can understand and solve adaptive controlproblems more easily, as the error models are independent of thespecific adaptive system
([email protected]) Tuesday 17-Feb-2015 3 / 27
Identification of a Vector Parameter in an Algebraic System
Problem:
y(t) = θ>u(t)
θ: unknowny(t), u(t): measured
Construct a parameter estimate θ
([email protected]) Tuesday 17-Feb-2015 4 / 27
Identification of a Vector Parameter in an Algebraic System
Difference the two output signals as
e = y − y
Express the parameter error as
θ = θ − θReduces to Error Model 1:
([email protected]) Tuesday 17-Feb-2015 5 / 27
Identification of a Vector Parameter in an Algebraic System
Error model 1:e = θ>u
Propose the candidate Lyapunov function:
V (θ) = θ>θ
Time differentiating
V = 2θ>˙θ
Adaptive law˙θ = −eu
Gives
V = −2θ>eu
= −2(θ>u)2
Thus V ≤ 0 ⇒ stability([email protected]) Tuesday 17-Feb-2015 6 / 27
Parameter Identification: Motivating Example
Transfer function of a DC motor:
V : Voltage inputω: Angular Velocity outputK,J,B: unknown physical parameters
Express the plant transfer function as:
ω
V=
K
Js+B=
a1s+ θ1
K,J,B unknown ⇒ a1, θ1 unknown
([email protected]) Tuesday 17-Feb-2015 7 / 27
Parameter Identification: Motivating Example
The differential equation describing the DC motor is
ω = −θ1ω + a1V
The DC motor is a first-order dynamical system
Recall: last time we assumed a1 was known
We looked at two procedures for identification
Error model 1Error model 3
Now we will consider both a1 and θ1 to be unknown
Will go through both procedures (error model 1 and 3) again
([email protected]) Tuesday 17-Feb-2015 8 / 27
Identification: First order plant, two unknown parameters
Plant Transfer Function:xpu
=a1
s+ θ1Express the plant transfer function as
xpu
=1
s+ θ1
=1
s+ θm· s+ θms+ θ1
=1
s+ θm· 1s+θ1s+θm
=1
s+ θm· 1θm−θms+θm
+ s+θ1s+θm
=1
s+ θm· 1
1− θm−θ1s+θm
where θm > 0 is a known positive parameter selected by control designer.([email protected]) Tuesday 17-Feb-2015 9 / 27
Identification: First order plant, two unknown parameters
Define θ , θm − θ1 and simplify this transfer function as
ω
u=
1
s+ θm· 1
1 + θs+θm
and realize this in the following block diagram representation, where weare using two states to represent a first order system.
([email protected]) Tuesday 17-Feb-2015 10 / 27
Identification: First order plant, two unknown parameters
Recall: when a1 known
y(t) ≡ xp(t)− a1φ(t) = θφ2(t)
⇒Error Model 1, Identify θ, a scalar
When a1 unknown:
y(t) ≡ xp(t) = a1φ(t) + θφ2(t) = θ>φ(t)
⇒Error Model 1, Identify θ, a vector([email protected]) Tuesday 17-Feb-2015 11 / 27
Identification: First order plant, two unknown parameters
y(t) = θ>φ(t)
We have just seen how to determine an update law for this system onslide 5: error model 1 with vector parameter
Now we will see an alternate method of identifying the two unknownparameters
([email protected]) Tuesday 17-Feb-2015 12 / 27
Identification: First order plant, two unknown parameters -An Alternate Method
The differential equation describing the plant is given by
xp = −θ1ω + a1u
Generate an estimate of the plant output as follows, using parameterestimates in place of the unknown parameters
˙xp = −θ1xp + a1u
Define the following output error and parameter errors
e = xp − xpθ1 = θ1 − θ1a1 = a1 − a1
([email protected]) Tuesday 17-Feb-2015 13 / 27
Identification: First order plant, two unknown parameters -An Alternate Method
The output error dynamics are given by
e = −θ1xp + a1u+ θ1xp − a1u
Add and subtract θ1xp
e = −θ1xp + θ1xp − θ1xp + a1u+ θ1xp − a1u= (θ1 − θ1)xp − θ1(xp − xp) + a1u
= −θ1e− θ1xp + a1u
= −θ1e+ θ>φ
where
θ =
[θ1a1
]and φ =
[−xpu
]([email protected]) Tuesday 17-Feb-2015 14 / 27
Identification: First order plant, two unknown parameters -An Alternate Method
Error Model 3:e = −θ1e+ θ>φ
We saw error model 3 last lecture for a system with one unknown
We will again determine an update law using a Lyapunov function
Choose a quadratic function V of the dominant errors in the system
V (e, θ) =1
2
(e2 + θ>θ
)Goal: Choose
˙θ so that V ≤ 0
([email protected]) Tuesday 17-Feb-2015 15 / 27
Identification: First order plant, two unknown parameters -An Alternate Method
Take the time derivative of V
V = ee+ θ˙θ
= −θ1e2 + θ>eφ+ θ>˙θ
Choose˙θ = −eφ
⇒ V = −θ1e2
Thus V ≤ 0⇒ stability.
([email protected]) Tuesday 17-Feb-2015 16 / 27
Adaptive Control
Everything we have seen thus far has been identification of unknownparameters
Now we introduce adaptive control: how to control systems withunknown parameters
([email protected]) Tuesday 17-Feb-2015 17 / 27
Adaptive Control of a First-Order Plant
Problem:
Plant:xp = apxp + kpu
Find u such that xp follows a desired command.
([email protected]) Tuesday 17-Feb-2015 18 / 27
A Model-reference Approach
xp: Output of a first-order system - can only follow ’smooth’ signals
Ensure xd is a ‘smooth’ signal essentially by filtering the desiredcommand
Pose the problem asxm = amxm + kmr
−2 0 2 4 60
0.2
0.4
0.6
0.8
1
Time [s]
xdxm
Set am = −5 and km = 5, say. Choose r so that xm ≈ xd([email protected]) Tuesday 17-Feb-2015 19 / 27
Statement of the problem
Choose u so that e(t)→ 0 as t→∞.ap unknown
kp unknown, but with known sign
([email protected]) Tuesday 17-Feb-2015 20 / 27
Certainty Equivalence Principle
Step 1: Algebraic Part: Find a solution to the problem when parametersare known.
Step 2: Analytic Part: Replace the unknown parameters by their estimates.Ensure stable and convergent behavior.
The use of the parameter estimates in place of the true parameters isknown as the certainty equivalence principle.
([email protected]) Tuesday 17-Feb-2015 21 / 27
Certainty Equivalence Principle- Step 1
Step 1: Algebraic Part: Propose the control law
u(t) = θcxp + kcr
and choose θc, kc so that closed-loop transfer function matches thereference model transfer function.
xp = apxp + kp(θcxp + kcr)
= (ap + kpθc)xp + kpkcr
Now compare this to the reference model equation
xm = amxm + kmr
Desired Parameters: θc = θ∗ and kc = k∗ must satisfy
ap + kpθ∗ = am and kpk
∗ = km
([email protected]) Tuesday 17-Feb-2015 22 / 27
Certainty Equivalence Principle- Step 1
Solving for the nominal or ideal parameters
θ∗ =am − apkp
and k∗ =kmkp
This is represented with the following block diagram
([email protected]) Tuesday 17-Feb-2015 23 / 27
Certainty Equivalence Principle - Step 2
Step 2: Analytic Part: Replace the unknown parameters by their estimates.Ensure stable and convergent behavior. From Step 1, we have
u(t) = θ∗xp + k∗r, θ∗ =am − apkp
, k∗ =kmkp
Replace θ∗ and k∗ by their estimates θ(t) and k(t).
u(t) = θ(t)xp + k(t)r
θ(t) =?? k(t) =??
([email protected]) Tuesday 17-Feb-2015 24 / 27
Certainty Equivalence Principle - Step 2
Adaptive control input:
u(t) = θ(t)xp + k(t)r
Define the parameter errors as
θ(t) = θ(t)− θ∗
k(t) = k(t)− k∗
Plug the control law into the plant equation:
xp = apxp + kpu(t)
= apxp + kp[θ(t)xp + k(t)r
]= apxp + kp
[θ(t)xp + θ∗xp + k(t)r + k∗r
]=
[ap + kpθ
∗]xp + kpθ(t)xp + kpk∗r + kpk(t)r
= amxp + kpθ(t)xp + kmr + kpk(t)r
([email protected]) Tuesday 17-Feb-2015 25 / 27
Certainty Equivalence Principle - Step 2
Reference Model:xm = amxm + kmr
Define the tracking error as
e = xp − xm
Error model:
e = ame+ kpθ(t)xp + kpk(t)r
= ame+ kpθT(t)ω
([email protected]) Tuesday 17-Feb-2015 26 / 27
ω =
xp
r
θ =
θ(t)
k(t)
Certainty Equivalence Principle - Step 2
e = ame+ kpθ>(t)ω
This is again error model 3!
([email protected]) Tuesday 17-Feb-2015 27 / 27