SE513: Systems Identification

Lecture 2
Introductory Examples
Dr. Samir Al-Amer

(c) 2007 Dr. Samir Al-Amer

Flow Chart for System Identification
Start
Apriori
Design of Experiment Knowledge

Perform Experiment and Collect data

Choose Model Structure

Choose method& Estimate parameters

Model Validation
No Yes
Model Accepted? STOP New data

(c) 2007 Dr. Samir Al-Amer

Lecture Outlines
O Important Concepts:
The System
The Model Structure
The Identification Method
Experimental Condition
O Basic Example
O Non-parametric Methods
O Parametric Methods

(c) 2007 Dr. Samir Al-Amer

Four Important Concepts in Identification

System Model
S Structure
M
Identification Experimental
Method Condition
J X
(c) 2007 Dr. Samir Al-Amer
 The System S
O The System S is a mathematical description of
the process
S is unknown in physical systems
No model can ever represent a real physical system
S is known in simulated systems
O The system S is specified by specifying
assumptions on the signals
O We do not need to know S to identify the system
O The system S is used for theoretical investigation
of behavior of different methods under different
conditions
(c) 2007 Dr. Samir Al-Amer
Model Structure M
O The model structure defines the set of all
candidate models parameterized by the
parameter vector

Example : y (t ) + a y (t 1) = b u (t 1) + (t )
(t ) : Error

O Candidate models are first order discrete time

models with two parameters a and b

(c) 2007 Dr. Samir Al-Amer

Identification Method
O The identification method is a procedure with
witch the parameter vector is obtained.
Model
structure Identification
(Model)
Data Method
O Large number of methods are available
O Choice depends on { model structure, data,
preference,.}

(c) 2007 Dr. Samir Al-Amer

Experimental Conditions X
O How to do the identification experiment?
O Major Decisions:
Selection of input signals
Selection of sample period
Selection of prefiltering
Take measurement while open loop/feedback

(c) 2007 Dr. Samir Al-Amer

 The System S is a mathematical
description of the process

The System S1
S is known in simulated systems

O We have two simulated systems

O S 1: y (t ) 0.8 y (t 1) = u (t 1) + e(t )
e : i.i.d. random variable
mean{e(t)} = 0, variance{e(t)} = 2

e(t)

1
1 0.8q 1
q 1
u(t) y(t)
1 0.8q 1
(c) 2007 Dr. Samir Al-Amer
 The System S is a mathematical
description of the process
S is known in simulated systems

The System S2
S 2: x(t ) 0.8 x(t 1) = u (t 1)
y (t ) = x(t ) + e(t )
mean{e(t)} = 0, variance{e(t)} = 2

e(t)

q 1 x(t)
u(t) y(t)
1 0.8q 1
(c) 2007 Dr. Samir Al-Amer
 Example 1
Non-parametric Model
O Data:Step response of S 1
O Model: Step Response Curve.

e(t)

1
1 0.8q 1

q 1
1 0.8q 1 y(t)
u(t):unit step

(c) 2007 Dr. Samir Al-Amer

 Example 2
Non-parametric Model
O Data:input and output of S 1
O Model: Weighting Sequence Curve.

e(t)

1
1 0.8q 1

q 1
1 0.8q 1 y(t)
u(t): white
noise
(c) 2007 Dr. Samir Al-Amer
Example 2
Non-parametric Model
O Data: u(t): white noise, y: response of S 1
O Model: Weighting Sequence Curve.
N k
1
N
y(t + k )u (t ) This is an approximate
h( k ) = t =1
N
formula for the
1
u
weighting sequence.
2
(t )
N t =1

(c) 2007 Dr. Samir Al-Amer

Weighting Sequence of S1

(c) 2007 Dr. Samir Al-Amer

True verses identified model (noise free step response)

Identified model
using correlation
method

True system

(c) 2007 Dr. Samir Al-Amer

Remarks on the obtained model
O The obtained model is not accurate
O The steady state gain is almost double
the true value
O The time constant and rise time are very
close to the true value

(c) 2007 Dr. Samir Al-Amer

Parametric Model
O True System S 1
O Model Structure M1: y (t ) + ay (t 1) = bu (t 1) + (t )

O Two unknown parameters = [a b ]T

O Why do we include in the model?

O is often called equation error

(c) 2007 Dr. Samir Al-Amer

Least Square Identification Method

O Determine the parameters of M such that the

sum of the square of the error is minimized.
O The loss function (to be minimized) is

N
V ( ) = 2 (t )
t =1
N
= [ y (t ) + ay (t 1) bu (t 1)]
2

t =1

O How do we minimize V( )?
(c) 2007 Dr. Samir Al-Amer
Least Square Identification Method

V ( ) = 0
a
The normal equations

V ( ) = 0
b

N N
N
y (t 1)
2
y (t 1)u (t 1)

a
y (t ) y (t 1)
N t =1 t =1
= Nt =1
y (t 1)u (t 1)
N
b
y (t )u (t 1)
u (t 1)
2

t =1 t =1

t =1

(c) 2007 Dr. Samir Al-Amer

Least Square Identification Method

O To find the unknown parameters, we need to

solve two equations in two unknowns

(c) 2007 Dr. Samir Al-Amer

Experimental Conditions X
O Selection of input signals
Step
Impulse
White Noise
PRBS (Pseudo Random Binary Sequence)
O Selection of sample period
O Selection of the number of samples N
O Selection of prefiltering
O Take measurement while open loop/feedback

(c) 2007 Dr. Samir Al-Amer

(c) 2007 Dr. Samir Al-Amer
Experimental Conditions X
O Different input sequences leads to
different identified models.
O Which one is better?
Step
Impulse
Sum of sinusoidal
PRBS (Pseudo Random Binary Sequence)

(c) 2007 Dr. Samir Al-Amer

Reading Material
O Chapter 2 (Soderstrom & Stoica)

(c) 2007 Dr. Samir Al-Amer