Model Reduction within Control Systems Design

Pepijn Wortelboer
PHILIPS OPTICAL STORAGE
January 2001

Introduction
Control problem formulation Need for order reduction algorithms Introduction of CD-player mechanism Standard reduction techniques: modal and balanced Closed-loop reduction Iterative low-order control design for CD-player Concluding remarks

Control problem formulation

Linear time-invariant model based standard plant formulation tuning by weighting function design

G
z

M K

Goal: simplest K achieving small z

z = I(G, M , K ) w

Need for order reduction algorithms

High-speed high-accuracy servo-mechanisms imply high-order models Highest frequency modes in model are not reliable High-order model is not suited for optimal control Controller order exceeds model order But order reduction not goal in itself Part of design

Introduction of CD-player mechanism

Rotating compact disc swing arm for radial tracking with optical pick-up unit (vertically suspended lens) for focussing mounting plate, disc, arm not infinitely stiff 2x2 servo control design problem

Radial control loop

wdist wtrack

No weights on input signals output weights

Ws to penalize low-frequent error e Wt to penalize high-frequent current i

Standard reduction techniques:

Modal
+ invariant for input & output + direct + no stability constraint

A diagonal

balanced
+ input-output significant

Gramians P & Q diag

Both suitable for truncation and singular perturbation Not optimal, how about closed-loop?

Closed-loop balancing
Transform

model state controller state

to get closed-loop system Gramians with

( TG
G

IM

( TK
K

G
M K

G
M K

closed-loop G-HSV closed-loop K-HSV

Closed-loop balanced reduction

Solve P, Q from Lyapunov equations of I(G,M,K) extract G-part: PG & QG find balancing transformation: T *QG T = T -1PG T -* = diag( HSVG ) apply T to G: Gbal =(T -1AG T, T -1BG , CG T ,DG ) truncate Gbal

Reduction in twin feedback configuration

G
M K
Tuning freedom: selection of order controller or model

( Gm = balRm I(Gh , M , K )

Separate weighting in the form of power spectra

G
z
M K
()
piecewise quadratic interpolation

( Gm = balRm z I (Gh , M , K ) w

Iterative low-order control design

Available: High order model G Weights in M First guess controller K

Model reduction =
Optimal control

I (Gh , M , K ) I (Gm , M , K ) I (Gh , M , K )

Controller reduction

= I (G , M , K r )

CD-player example
Model G120 Ws order 1, Wt order 2 First guess order-3 controller:

Sequence of model and controller reduction

G 120
order a a
Models (G)

b
Controllers (K)

G20
d

K23
e c c

Order reduction Controller synthesis

G 14

K17
f f

K11

K9

K3

Closed-loop evaluation
23 = 913.76
open-loop -G *K
b b

17 = 913.77
controller K

K23 b

60 40 20 [dB] [dB] 0 -20 -40 -60 -80 2 10

3 4

20 15 10 5 0 -5 -10 -15 2 10
3

K17

Gb Kb
10 10

Kb
10 10
4

sensitivity 20 10 0 [dB] [dB] -10 -20 -30 -40 -50 2 10

3 4

complementary sensitivity 20 0 -20 -40

No visual difference

Sensitivity
-60 -80 2 10 10 Frequency [Hz] 10

Complementary Sensitivity
10 Frequency [Hz]
3

Model reduction OK
10
4

HSV and performance measures

10
-3

step aa 2500

step cc

10

-4

gamma & estimates

delta & estimates

HSV(G)

2000 1500 1000 500 0

HSV(K)
4 6 8 10 1 2 14 16 18 20 reduced order controller step ff 22

10

-5

10

-6

10

20 30 reduced order model step dd

40

10

2500

10

-2

gamma & estimates

delta & estimates

HSV(G)

2000 1500 1000 500 0

HSV(K)
4 6 8 10 12 14 reduced order controller 16

10

-4

10

-6

10 15 reduced order model

20

Models
40 20

G120
0 -20 G120 G120-G20 G20-G14

-40

-60

G120-G20

-80

-100

G20-G14
-120 2 10 10
3

10

20 K23 K9 K3 15

Controllers
K9 K23

10

K3
-5 -10

-15

-20 2 10

10

10

20

Direct controller reduction with manually shaped frequency weights

K23 Kreduced error weighting
40

Plain balanced reduction

15 10

23
0 -5

12

-10

-15 2 10

10

10

40

30

30

20

10

Interval weighted balanced reduction 23 to 11

20

10

Interval weighted balanced reduction 23 to 13

-10

-10

-20

-20

-30

-30

-40 2 10

10 max. sv of G23, G11, and difference

10

-40 2 10

10 max. sv of G23, G13, and difference

10

Recommendations from Experience:

See order reduction as a tool for control design Use robust and fast algorithms Exploit specific structure of model Apply iterative scheme (with frequency tuning) Evaluate results in more than one measure (graphics) Choose odd or even order Do not rely on a priori HSV-based error bounds Customize and automate in later stage
P.M.R. Wortelboer, M. Steinbuch and O.H. Bosgra Iterative Model and Controller Reduction using Closed-loop Balancing, with application to a Compact Disc Mechanism Int. J. Robust Nonlinear Control 9, 123-142 (1999)

Further reading: