Trend and System Identification With Orthogonal Basis Functionweb.abo.fi/fak/tkf/at/ose/doc/Pres_15112013/Amir Shirdel.pdf · Shirdel : Trend and System Identification With Orthogonal

Trend and System Identification With Orthogonal Basis Function

OSE SEMINAR 2013

ÄMIR SHIRDEL

CENTER OF EXCELLENCE IN

OPTIMIZATION AND SYSTEMS ENGINEERING

AT ÅBO AKADEMI UNIVERSITY

ÅBO NOVEMBER 15 2013

Background

– System identification is difficult when process measurements are

corrupted by structured disturbances, such as trends, outliers, level

shifts

– Standard approach: removal by data preprocessing but difficult to

separate between the effects of known system inputs and unknown

disturbances (trends, etc.)

– Orthonormal basis function models are categorized as output-error

(ballistic simulation) models

2|N

Background

Amir Shirdel: Trend and System Identification With Orthogonal Basis Function Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University

2|22

Present contribution

– Identification of system model parameters and disturbances

simultaneously

– Sparse optimization used in system identification problem

– Applying the method on simulated and real example

3|N


Amir Shirdel: Trend and System Identification With Orthogonal Basis Function Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University

3|22

4

4 4


Amir Shirdel : Trend and System Identification With Orthogonal Basis Function Center of Excellence in Optimization and Systems Engineering at Åbo Akademi University

4|22


SYSTEN

(model) U Y

0 100 200 300 400 500 600 700 800 900 1000-10

-8

-6

-4

-2

0

2

4

6

8

d

0 100 200 300 400 500 600 700 800 900 1000-25

-20

-15

-10

-5

0

5

10

15

20

0 100 200 300 400 500 600 700 800 900 1000-4

-3

-2

-1

0

1

2

3

4

5

Orthogonal basis functions has some advantages:

• The corresponding approximation (representation) has simple and direct

solution.

• It corresponds to allpass filters which is robust to implement and use in

numerical computation.

• It is popular because a few parameters can describe the system.

• It is a kind of output-error model, and can be insensitive to noise.

5 5

Orthogonal basis function (Fixed-pole model)


5|22

Orthogonal basis function

6 6 6

Orthogonal basis function (Fixed-pole model)


6|22

Orthogonal basis function

FIR network:

Laguerre function:

Kautz function:

7

Model:

where

The parameters can be estimated using standard least squares method:

7 7

System Identification


7|22

System identification

2)ˆ)(( θ(k)ky T

d(k)θ(k)y(k) T

System u

d

y

T

n

T

n aaakukukuk ],...,,[,)](),...,(),([)( 2121

)()()(or)()()( kuqkukuqLku nnnn

8 8 8

Problem formulation


8|22

Problem formulation

We consider the linear system model:

It is assumed the measured output is given by

where d(k) is a structured disturbance:

• outlier signal,

• level shifts,

• piecewise constant trends

)()()( kdkyky L

),(...)()()( 2211 kuakuakuaky nnL

9 9 9

Disturbance models


9|22

Disturbance models

Sequence of outliers:

Level shifts:

Sequence of trends:

10 10 10

Sparse representation of disturbance


10|22

Sparse representation of disturbance

For the structured disturbances, the vectors are sparse,

where and , depends on disturbances

Outliers:

Level shifts:

Trends:

dDi

ID 0

iD

11

Identification by sparse optimization:

subject to

where

This is an intractable combinatorial optimization problem.

Instead we use and solve the convex problem

11 11 11 11

Sparse optimization approach


11|22

Sparse optimization approach

1

2ˆ,ˆ

ˆ))(ˆ)((min dDλkyky i

kd

2ˆ,ˆ

))(ˆ)((min k

dkyky

elements nonzero ofnumber 0

MdDi 0

relaxation1 l

12

Algorithm


12|22

Algorithm

1. Make the basis function expansion based on given prior knowledge of system (pole and system order).

2. Solution of sparse optimization problem by iterative reweighting:

Minimize the weighted cost to give the estimates

where i = 0, 1 or 2 (user selected).

3. Calculate new weights W and go to step 2.

4. Continue until convergence.

5. Use model order reduction to get lower order model.

1

2ˆ,ˆ

ˆ))(ˆˆ)((min dDWλkdky ii

k

T

d

We apply the proposed identification detrending method to the ARX model

with parameter vector

given by

u(k) and e(k) are normally distributed signals with variances 1 and 0.1, and

d(k) is unknown structured disturbance ,

For making the Kautz basis function we used N=4 as order of system and pole

is

13 13

Example1

13|22

Example


4.07.0

14

Example1

14|22

Example


0.9340ˆ 9125.0 ˆ RMSE :without

5.2937RMSE 0.7697- 1.3465 2.8084 1.5214 ˆ

0.9102RMSE 0.8412- 1.3152 2.7305 1.5428 ˆ

LS

LS

RMSEd(k)

0 100 200 300 400 500 600 700 800 900 1000 -20

-10

0

10

20

K data points

Output

Output (model)

Output (LS)

0 100 200 300 400 500 600 700 800 900 1000

-5

0

5

10

K data points

Disturbance (Identified)

Disturbance

15

Example1

15|22

Example


0 1 2 3 4 50

1

2

3

4

5

6

7

8Hankel Singular Values (State Contributions)

State

Sta

te E

nerg

y

Stable modes

16

Example1

16|22

Example


0.9118RMSE 0.4933- 0.6946- ˆ

0.9102RMSE 0.8412- 1.3152 2.7305 1.5428 ˆ

R

0 100 200 300 400 500 600 700 800 900 1000-20

-10

0

10

K data points

Output (model)

Output (Reduced model)

0 100 200 300 400 500 600 700 800 900 1000

-5

0

5

10

K data points

Disturbance (Identified)

Disturbance

17 17

Example2: Pilot-scale distillation column data

17|22

Example


0 200 400 600 800 1000 1200 1400 1600 1800 2000 91

91.5

K data points

Top column output

0 200 400 600 800 1000 1200 1400 1600 1800 2000 0

2

K data points

Bottom column output

0 200 400 600 800 1000 1200 1400 1600 1800 2000 156

158

160

162

K data points

Reflux (L)

0 200 400 600 800 1000 1200 1400 1600 1800 2000 87

88

89

90

K data points

Reboiling flow (V)

1

18 18

Example2: Identification results

18|22

Example


0 200 400 600 800 1000 1200 1400 1600 1800 2000-1

-0.5

0

0.5

1

1.5

K data points

Output (real)

Output+dd (model)

0 200 400 600 800 1000 1200 1400 1600 1800 2000-0.2

-0.1

0

0.1

0.2

0.3

K data points

Disturbance (identified)

0.0933RMSE 0.0456- 0.0353- ˆ

0.0826 RMSE 12) ...(total 0.0147- 0.0338- 0.0285- 0.0022- ˆ

0.0643RMSE 12) (total ... 0.0151- 0.0346- 0.0280- 0.0031- ˆ

RED

LS

19 19

Example2: Hankel singular value for reduction

19|22

Example


0 2 4 6 8 10 120

0.1

0.2

0.3

0.4

0.5

0.6

0.7Hankel Singular Values (State Contributions)

State

Sta

te E

nerg

y

Stable modes

20 20

Example2: Validation data with estimation of d(k)

20|22

Example


0.0891RMSE 0.0456- 0.0353- ˆ RED

0 200 400 600 800 1000 1200-1

0

1

Output (real)

Output(Test Model+Disturbance)

0 200 400 600 800 1000 1200

-0.2

0

0.2

Disturbance

0 200 400 600 800 1000 1200

-0.5

0

0.5

0 200 400 600 800 1000 1200-1

0

1

Summary:

• Presented a method for identification of linear systems in the presence of

structured disturbances (outliers, level shifts and trends) by using sparse

optimization and orthogonal basis function as the system model

• Gives acceptable results for simulated example

• Gives acceptable results for distillation column example

• The orthogonal basis model improves the robustness and more insensitive

to noise

Future work:

• Nonlinear system identification and trends and more general disturbances

21 21

Discussion and future work

21|22

Discussion and future work


22 22

Thank you for your attention!

Questions?

22 22

22|22


Documents

Trend and System Identification With Orthogonal Basis Functionweb.abo.fi/fak/tkf/at/ose/doc/Pres_15112013/Amir Shirdel.pdf · Shirdel : Trend and System Identification With Orthogonal