Chapter 3 System ID

1. Scalar case, one unknown parameter, intuitive method 1 2. Scalar case, one unknown parameter, intuitive method 2 3. Scalar case, one unknown parameter, un-normalized 4. Scalar case, one unknown parameter, normalized 5. First order linear system-scalar case, un-normalized 6. First order linear system-vector case, un-normalized 7. First order linear system-vector case, normalized 8. SPR Lyapunov approach, normalized 9. Gradient method, normalized 10. Least squares, normalized

Chapter 3 System ID for adaptive control

1. Scalar case, one unknown parameter, intuitive method 1 2. Scalar case, one unknown parameter, intuitive method 2 3. Scalar case, one unknown parameter, un-normalized method 4. Scalar case, one unknown parameter, normalized method 5. First order linear system-scalar case, un-normalized 6. First order linear system-vector case, un-normalized 7. First order linear system-vector case, normalized 8. SPR Lyapunov approach, normalized 9. Gradient method, normalized 10. Least squares, normalized

1. Scalar case, one unknown parameter, intuitive method 1 System : y (t) = .u (t) Where : y (t) u (t) = output = input = unknown constant Assume that y (t) and u (t) are measurable of . Solution : Get the 1/0 pair {u (t) , y (t)} over an interval [ , some t [ , + T] and choose as Remark : u (t) maybe very close to zero = very big! Unbounded Remark : It would be worse if the system is subjected to noise The system would result in a large spread in the value of at different t [ , + T] Example : y (t) = . u (t) + d Review y (t) = if u 0 the problem is = of u 0 ; may have measuring error/noise by u 2. ID Scalar case, one unknown parameter, 2nd intuitive method Lets define an estimator with t [ , + T] system : y = . u(t) . u(t) , [ , + T] + T] so that u (t) 0 for t to and it is desired to obtain an estimate

The solution maybe considered as one that minimizes the performance index (p1) : Where [ [ ] [ ] ]

Its minimization is evaluated as : [ ][ ]

for u (t) = 0 , t [ , + T]

Remark : as in the previous method, can only be calculated once if no noise present Remark : If noise shows up, it might not gine true of Remark : Increasing of I will improve To anticipate it, we meed RLS : recursive lee square! Because its impossible to increment the function to another time interval 3. ID Scalar case, one unknown parameter, unvormalized (the system) System y(t) = Where u . is an unknown constant and u,y are measurable from the measured data

We would like to determine Define = = - ( =- Define J( = = J( ( )u

lets consider the estimated output :

output error

performance )

J( is pointuise of J

is convex over

and hence the minimization

of J is global convex optimization problem ECX is said to be convex if = = (

[0,1] such that ]

* ( ) ) +

Stability of the estimator above can be proved as : Let V ( 0 (proef) = < as t = Check = If then 0 as t = y= u Because of : V = If When will as t = = ? as t = J= (Barbalats lemma) ( ) 0 V(x) is p.d if V (x) = 0 V (x) > 0 parameter error is bounded ( Check x=0 x

A necessary and sufficient condition is that u satisfies the PE condition and for some PE = Persistent Exatation U should be rich enough to have parameter convergence 4. ID Scalar case, one unknown parameter, normalized System y(t) = Where is an unknown constant and u,y are measurable u Since now u could be unbounded, we may not set Because it would give an unbounded Let us divide the system equation by the normalizing signal m > 0 To have Where and , and m2 = 1 +

The signal ns is chose so that e.g. pick ns = u m2 = 1 +

u2

Consider the estimate Let = = ( J( = = = ( ( ) ) normalized output error ( ) )

= = = Let V = = =

) =m>0

Using the gradient method * ( ) + ) +

* ( = -

<

= if then as t = as t =

For parameter convergence, me need PE Persistent excitation ID of 1st order linear systems (unnormalized) Plant : Where Xp R is the output of the plant and u R is the bounded input

Ap and Kp are unknown constants, but me know that ap < 0 The problem is to determine ap and kp from u and xp

Here, we are going to solve this problem with 2 estimators :

Parallel estimator output error method only for scalar system (not for vector) Series parallel estimator equation error method for scalar and vector systems

Method 1 : - output error method Select the parallel estimator

( ) ) ( ( ) ( ) )u

- error dynamics ,

Where e =

(The output error dynamics is driven by the output error) And parameter error

Y (t) = u

st

not discussed yet

5. ID of 1 order linear system (unnormalized) Plant =

Where Xp

IR , u

R are measurable

Ap and Kp are unknown constants, ap < 0

Parallel estimator output error method (only for scalar) Series parallel estimator equation error method (also for method)

Method : - output error method Select the parallel estimator

( )

( ) ) ( ( )u )

- error dynamics

Define a Lyapunov function candidate

Take the time derivate of V along the trajectory of the errors dynamics, we have : ( Bounded )

Pick

[ ]

Stable LTI System

Xp

BI Bounded Input

BIBO Stable

BO Bounded Output

 We need PE for parameter Convergence

Summary : For parameter Convergence, we need PE

For parameter Convergence, we need PE

Method 2 : - Equation error method Select the series-paralel estimator : ( ( )

( ( )

[ )

= [ t<0

] )

Pick

. We need PE for convergence of the parameters. ID VECTOR CASE UN-FORMALIZED Plant = Where Assume that is Hurwitz and u is bonded. We would like to find and , based on the information of Method : Output error method Define the parallel estimator : The error dynamics can be as : Pick the Lyapinov function candidate : ( ) =( ( = ( ( ) ) ) ( ) )

and u.

? Although is known to be stable, its elements are not known and hence a matrix p > 0 cant be found s.t. with Q > 0. This approach doesnt work for veetos case but its fine for scalar case. Method 2 : Equation error method Define the series-parallel estimator ( ) Where and and is known Let V = =( ) ( ( ( ) ) ) ( ) ( ) [ ]

( Method 2 : =-Q is stable

) s, t, (

) )

) ) ( (

) )

( ( Pick ( ( ) ) ( ( ) )

( )

Update laus

( because q is a constant

) ( )

We need PE for parameter converage! What about u ( Plant : Let ) ( )

7. ID-1st order systems, normalized u unknown constans +

Define e = x - Where s is the laplace variable Ns is to be designed s.t. and with and ns =

(s-am)e = (S-am) . (x - ) e. - - ) am (x - ) e. = (ax bu) (am + ( -am) x + = am (x - ) + (a- x + (b- )u = (am = (am (x - ) + (a- x + (b- )u am (x - ) - e. e + (a- x + (b- )u e+ + Let V= (am [ (am xe ue ( ( ) - ] ) )

E,

For some n > 0 =

We can not conclude asympt stability of e e.ns <proof> :v ( Q.E.D )

e. m = < <proof) : =

<proof) :

Whats the effect of ns ? (check) : Let ns = 1+

Then from the equation = (am e=

e+ +

for steady state

=
= =

Where

Linear Parametric Model (LPM) SISO Plant : Y = CT.X Where X E Rn , X(o) = Xo , y ,u ER are measurable

= If Where , then we have : Z(s) = R(s) = R(s) y = Z(s) . u

=[ =

In general, we represent a linear parametric model as : Y = G(s) .

Review e = x-
updated laws : Positive Linear System SPR : Strictly positive Real

Positive linear systems are linear systems with positive real T.F.s System :

(n-th order SISO linear system) Def : h(s) is positive real (PR) if Re [h(s)] 0 Re [s] 0

h(s) is strictly positive real (SPR) if h (s- ) is PR for some > 0 e.q. h(s) = h( )= = If Re [h(s)] = Then h(s) is PR For SPR : Pick h(s - ) = is P.R
( )

> 0, S=
[ ]

[ ]

is SPR Thm : h(s) is SPR iff h(s) is strictly stable, and Re [h(j )] > 0 Some useful necessary conditions to check SPR : 1. h(s) strictly stable e.g h(s) = is not SPR

because the evalue is not in the left hand plane 2. h(s) has relative degree 0 or 1 e.g h(s) = is not SPR

3. h(s) is strictly minimum phase e.g h(s) = is not SPR

4. the ngguist plot of h(j ) lies entirely in RHP (Right Hand - Plane) 5. phase shift in response to sinusoidal input is always less than 900 the kalman-yakuborich lemma (positive real lemma) lemma : system : y= (A,B) controllable The T.F. h(s) = Iff is SPR

p > 0 and Q > 0 such that (s.t)

PB = C Controllable stabilizable Uncontrollable PC stabilizable Co stabilizable stabilizable stabilizable stabilizable f is smooth is continuous uncontrollable not stabilizable

p.c. stabilizable

piece wise continuous

modified k-y-lemma kaliman-Yakubovich-Meyer Lemma given x if the T.F. vector q and p = PT > 0 s.t 0 A asymptotically stable, symmetric p.d. matrix L is SPR, then

PB = C + 8. ID General SISO plant SPR Lyapunov Approach, normalized Plant : y = G(s) . ,y

Choose L(s) s.t. L-1(s) is a proper stable T.F. and G(s) . L(s) is a proper SPR T.F y = G (s) . L(s) . = G (s) . L(s) . e.q : y =
= = G(s). [ ] * +

Let L = s+2 Y = G L L-1

L-1 =

] *

][

An estimator can be constructed as

Case 1 : G(s) is minimum phase Pick L (s) = G-1(s) s.t . GL = 1

Define
(

( ) ( )

etc Case 2 :

G(s) is not required to be minimum phase ( Define the normalized error )

(
GL

)
e

Realization to state-space function

( )

)]

[ ( )]

 ( )

1D Gradient Method - normalized System:

Two approaches are introduced below based on different cost-function Case 1 :Instantaneous cost function ( ) (

Convexity of ( ) guarantees the existense of a unique global Minimum devine by : ( ) ( )

(1) Non-recursive algorithm 1 : ( ) (

( ) ( )

(2) Non recursive alogarithm 2:

[ [

] ] ( )

Adapptive control of linear-time-varying system (phD Thesis) Joannou and Tsaklis

(3) Recursive Algorithm ( ) ( )

( ) ( )

March 30, 2006

( )

*it means : ( )

}* to make

Case 2 : Integral Cost Function ( )

( )

( )

( ) [

( (

Homogeneous

Note that R(t) and Q(t) are exactly solutions to the following two Non-homogenous D.E. (differential equation) respectively { }