37
SYSTEMS SYSTEMS Identification Identification Ali Karimpour Assistant Professor Ferdowsi University of Mashhad [email protected] Reference: “System Identification Theory For The User” Lennart Ljung(1999)

SYSTEMS Identification

  • Upload
    kalani

  • View
    28

  • Download
    0

Embed Size (px)

DESCRIPTION

SYSTEMS Identification. Ali Karimpour Assistant Professor Ferdowsi University of Mashhad [email protected]. Reference: “System Identification Theory For The User” Lennart Ljung(1999). Lecture 6. Nonparametric Time and Frequency domain methods. Topics to be covered include : - PowerPoint PPT Presentation

Citation preview

Page 1: SYSTEMS Identification

SYSTEMSSYSTEMSIdentificationIdentification

Ali Karimpour

Assistant Professor

Ferdowsi University of Mashhad

[email protected]

Reference: “System Identification Theory For The User” Lennart Ljung(1999)

Page 2: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

2

Lecture 6

Nonparametric Time and Frequency domain methods

Topics to be covered include:

Transient-Response Analysis and Correlation Analysis

Frequency Response Analysis

Fourier Analysis

Spectral Analysis

Estimating the Disturbance Spectrum

Page 3: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

3

Nonparametric Time and Frequency domain methods

Topics to be covered include:

Transient-Response Analysis and Correlation Analysis

Frequency Response Analysis

Fourier Analysis

Spectral Analysis

Estimating the Disturbance Spectrum

Page 4: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

4

Transient-Response Analysis and Correlation Analysis

Impulse Response Analysis

Suppose

Let the input as

Then the output will be

So we have

The error is:

This simple idea is impulse Response Analysis.

Its basic weakness is that many physical processes do not allow pulse input so the the error be small, moreover that input make the system exhibit nonlinear effects.

)()()()( 0 tvtuqgty

00

0)(

t

ttu

)()()( 0 tvtgty

)(

)(ˆty

tg

)(tv

Page 5: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

5

Transient-Response Analysis and Correlation Analysis

Step Response Analysis

Suppose

Let the input as

Then the output will be

So we have

The error is:

)()()()( 0 tvtuqgty

00

0)(

t

ttu

)()()(1

0 tvkgtyt

k

)1()(

)(ˆ

tyty

tg

)1()( tvtv

It would suffer from large error in most practical applications.

It is acceptable for delay time, static gain, dominating time constant.

Page 6: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

6

Transient-Response Analysis and Correlation Analysis

Correlation Analysis

Suppose

If the input is a quasi stationary sequence with

Then we have

• If the input is white noise

Then if the input is white noise

An estimate of the impulse response is thus obtained from

)()()()( 0 tvtuqgty

)()()( uRtutuE

0)( uR

)(

)(0yuR

g

)()()()(1

0 tvktukgtyk

operation loop-open 0)()( tvtuE

1

0 )()()()()(k

uyu kRkgRtutyE

)(

)(0yuR

g

N

t

Nyu tuky

NR

)()(

1)(

• If the input is not white noise

Page 7: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

7

Transient-Response Analysis and Correlation Analysis

Correlation Analysis

Suppose

If the input is a quasi stationary sequence with

Then we have

• If the input is white noise

Then if the input is not white noise

Choose the input so that (I) and (II) become easy to solve.

)()()()( 0 tvtuqgty

)()()( uRtutuE

)()()(1

)(ˆ ItutuN

RN

t

Nu

)()()()(1

0 tvktukgtyk

operation loop-open 0)()( tvtuE

1

0 )()()()()(k

uyu kRkgRtutyE

• If the input is not white noise

)()(ˆ)(ˆ)(ˆ1

IIkRkgRM

k

Nu

Nyu

Page 8: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

8

Transient-Response Analysis and Correlation Analysis

Exercise 1:

Suppose)()4(1.0)3(25.0)2(5.0)1()( tvtututututy

and v(t) is normal noise such that

)2,0()( Ntv

a) Derive with Impulse Response Analysis let u(t)<3)(ˆ tg

)(ˆ tgb) Derive with Step Response Analysis let u(t)<3

)(ˆ tgc) Derive with correlation analysis with white noise input let u(t)<3

)(ˆ tgd) Derive with correlation analysis with non white noise input u(t)<3

Page 9: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

9

Nonparametric Time and Frequency domain methods

Topics to be covered include:

Transient-Response Analysis and Correlation Analysis

Frequency Response Analysis

Fourier Analysis

Spectral Analysis

Estimating the Disturbance Spectrum

Page 10: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

10

Frequency-Response Analysis

Sine Wave testing

Suppose

Let the input as

Then the output will be

This is known as frequency analysis and is a simple method for obtaining detailed information about a linear system.

Bode plot of the system can be obtained easily.

One may concentrate the effort to the interesting frequency ranges.

Many industrial processes do not admit sinusoidal inputs in normal operation.

Long experimentation periods.

)()()()( 0 tvtuqgty

...,2,1,0cos)( tttu

parttransient )(cos)()( 0 tvtegty i ??

)()( 0 iegt

Page 11: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

11

Frequency-Response Analysis

Sine Wave testing

...,2,1,0cos)( tttu parttransient )(cos)()( 0 tvtegty i ??

)()( 0 iegt

Frequency analysis by the correlation method.

If v(t) does not contain a pure periodic component of frequency ω .

0

Define ,sin)(1

)(,cos)(1

)(11

ttyN

NIttyN

NIN

ts

N

tc

0

0 0

2/

)()()(ˆ

22

NINI

eg csiN

)(

)()(ˆ

NI

NIarctgeg

c

siN

Page 12: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

12

Transient-Response Analysis and Correlation Analysis

Exercise 2:

Suppose)()4(1.0)1(25.0)2(5.0)1()( tvtututututy

and v(t) is normal noise such that

)2,0()( Ntv

a) Derive and discuss about the value of N.)(ˆ iN eg

Page 13: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

13

Nonparametric Time and Frequency domain methods

Topics to be covered include:

Transient-Response Analysis and Correlation Analysis

Frequency Response Analysis

Fourier Analysis

Spectral Analysis

Estimating the Disturbance Spectrum

Page 14: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

14

Fourier Analysis

In a linear system different frequencies pass through the system independently.

So extend the frequency analysis estimate to the case of multifrequency inputs.

Empirical Transfer-Function Estimate

Properties of ETFE

Claim:

Remember:

)(

)()(ˆ

N

NiN U

Yeg

Page 15: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

15

Fourier Analysis

Properties of ETFE

Claim:

Remember:

Now suppose

Let

By above claim

)()()()( 0 tvtuqgty

N

t

tiN etv

NV

1

)(1

)(

)(

)(

)(

)()()(ˆ 0

N

N

N

NiiN U

V

U

Regeg

Page 16: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

16

Fourier Analysis

Properties of ETFE

Since v(t) is assumed zero mean

So

)()()()( 0 tvtuqgty

N

t

tiN etv

NV

1

)(1

)(

)(

)(

)(

)()()(ˆ 0

N

N

N

NiiN U

V

U

Regeg

0)(NEV

)(

)()()(ˆ 0

N

NiiN U

RegegE

Page 17: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

17

Fourier AnalysisProperties of ETFE

So

N

t

tiN etv

NV

1

)(1

)(

)(

)(

)(

)()()(ˆ 0

N

N

N

NiiN U

V

U

Regeg

)(

)()()(ˆ 0

N

NiiN U

RegegE

)()(ˆ)()(ˆ 00 ii

Nii

N egegegegE )()()()( 0 tvtuqgty

Page 18: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

18

Nonparametric Time and Frequency domain methods

Topics to be covered include:

Transient-Response Analysis and Correlation Analysis

Frequency Response Analysis

Fourier Analysis

Spectral Analysis

Estimating the Disturbance Spectrum

Page 19: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

19

Spectral Analysis

Smoothing the ETFEThe true transfer function is a smooth function.

If the frequrncy distance 2π/N is small compared to how quickly

Changes then

)(0jeG

Are uncorrelated and if we assume

To be constant over the interval

)(0jeG

Page 20: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

20

For large N we have

If transfer function is not constant

Spectral Analysis

Page 21: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

21

If noise spectrum is not known and don’t change very much over frequency intervals

Then

Spectral Analysis

Page 22: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

22

Connection with the Blackman-Tukey Proceture

If N

If

Spectral Analysis

Page 23: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

23

If noise spectrum don’t change much over frequency intervals

dUW NNu

2

00 )()()(ˆ

similarly

Spectral Analysis

Page 24: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

24

The fourier cofficients for periodogram 2

)(NU

The fourier cofficients of the function W2

Spectral Analysis

Page 25: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

25

smooth function is chosen so that its fourier coefficients vanish

for

Weighting Function: The Frequency Window

Spectral Analysis

Page 26: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

26

Spectral Analysis

Page 27: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

27

Asymptotic Properties of the Smoothed Estimate

1-Bias

2-Variance

Here

Spectral Analysis

Page 28: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

28

5/1

5/122

.)(

)()(4N

W

RM

v

u

opt

Value of the width parameter that minimizes the MSE is

The optimal choice of width parameter leads to a MSE error that decays like

5/4. NCMSE

Spectral Analysis

Page 29: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

29

Page 30: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

30

Spectral Analysis

Smoothing the ETFEExercise 3:

)()2(5.0)1()2(7.0)1(5.1)( tetututytyty

Page 31: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

31

Another Way of Smoothing the ETFE

Spectral Analysis

The ETFEs obtained over different data sets will also provide uncorrelated estimates, and another approach would be to form averages over these. Split the data set into M batches, each containing R data (N=R . M). Then form the ETFE corresponding to the kth batch:

The estimate can then be formed as a direct average

ikR eG )(ˆ Mk ,...2,1

M

k

ikR

iN eG

MeG

1

)(ˆ1ˆ

Page 32: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

32

Or one that is weighted according to the inverse variances:

with

M

k

kR

M

k

ikR

kR

iN

eGeG

1

)(

1

)()(

)(

ˆ)(ˆ

2)()( )( kR

kR U

Page 33: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

Estimating The Disturbance Spectrum

)()()()( 0 tvtuqgty Suppose

dUW NNv

2

00 )()()(ˆ 1-Bias

)1())(()()(2

1)()(ˆ

1 NOCOME Nvv

Nv

2-Variance)1()(

)()(ˆ 2

NON

WVar Nv

Nv ,0

Page 34: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

34

The Residual Spectrum

)()(ˆ)()(ˆ tuqGtytv N

dUeGYW Ni

NNNv

2

0 )()(ˆ)()()(ˆ

Page 35: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

35

We have

)(ˆ

)(ˆ)(ˆ)(ˆ

2

Nu

NyuN

yNv

Coherency Spectrum:

Denote

)(ˆ)(ˆ

)(ˆ)(ˆ

2

Nu

Ny

NyuN

yuk

Then

2)(ˆ1)(ˆ)(ˆ N

yuNy

Nv k

Page 36: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

36

Nonparametric Time and Frequency domain methods

Topics to be covered include:

Transient-Response Analysis and Correlation Analysis

Frequency Response Analysis

Fourier Analysis

Spectral Analysis

Estimating the Disturbance Spectrum

Page 37: SYSTEMS Identification

lecture 6

Ali Karimpour Nov 2010

37

Estimating the Disturbance Spectrum

Properties of ETFE