SYSTEMS

Identification

Ali Karimpour
Assistant Professor
Ferdowsi University of Mashhad
Reference: System Identification Theory For The User Lennart Ljung(1999)
Practical Issues of System Identification Lennart Ljung (2007)
Perspectives on System Identification Lennart Ljung (2009)

lecture 1

Lecture 1
Perspective on System Identification
Topics to be covered include:

System Identification.

Place System Identification on the global map. Who are our neighbors in this
part of universe?

Discuss some open areas in System Identification.

Ali Karimpour Sep 2010

lecture 1

System Identification
System Identification: The art and science of building
mathematical models of dynamic systems from observed
input-output data.
System Identification is look for sustainable description by
proper decision on:
Model
complexity

Information
contents in the
data

Effective
Validation

Ali Karimpour Sep 2010

lecture 1

Dynamic systems
System: An object in which variables of different
kinds interact and produce observable signals.
Stimuli: External signals that affects system.
Dynamic System: A system that the current
output value depends not only on the current
external stimuli but also on their earlier value.
Time series: A dynamic system whose external
stimuli are not observed.
4

Ali Karimpour Sep 2010

lecture 1

Dynamic systems
Stimuli
Input

Disturbance

It can be manipulated
by the observer.

It can not be manipulated

by the observer.
Measured

Measured
disturbance
w
u
Input

Unmeasured

Unmeasured
disturbance

Dynamic
system

y
Output

Ali Karimpour Sep 2010

lecture 1

A solar heated house

Solar
radiation
w
u
Pump
velocity

Wind, outdoor
temperature

Dynamic
system

y
Storage
temperature

Ali Karimpour Sep 2010

lecture 1

Speech generation

chord, vibaration
airflow

Dynamic
system

y
Sound

Time series: A dynamic system whose external

stimuli are not observed.
7

Ali Karimpour Sep 2010

lecture 1

Models
Model: Relationship among observed signals.
1- Mental models
Model
types

2- Graphical models
3- Mathematical (analytical) models
4- Software models
Split up system into subsystems,

1- Modeling

Building
models

Joined subsystems mathematically,

Does not necessarily involve any experimentation
on the actual system.
It is directly based on experimentation.

2- System identification
3- Combined

Input and output signals from the

system are recorded.

Ali Karimpour Sep 2010

lecture 1

The fiction of a true model

Ali Karimpour Sep 2010

lecture 1

The Core
The Core: The core of estimating models is statistical theory.
Model: m
True Description: S
Model Class: M
Complexity (Flexibility): C
Information: Z
Estimation
Validation
Model Fit: F(m,Z)

10

Ali Karimpour Sep 2010

lecture 1

Estimation
A template problem: Curve fitting
Squeeze out the
relevant information
in data.

No more satisfaction

All data contains signal and noise.

11

Ali Karimpour Sep 2010

lecture 1

Estimation
The simplest explanation is usually the correct one. So the
conceptual process for estimation is:

Fit measure

Complexity measure

good agreement with data

Not too complex

m is a random variable since of irrelevant part of data (noise).

12

Ali Karimpour Sep 2010

lecture 1

The System Identification Problem

1- Select an input signal to apply to the process.
2- Collect the corresponding output data.
3- Scrutinize the corresponding output data to find out if some
preprocessing
4- Specify a model structure.
5- Find the best model in this structure.
6- Evaluate the property of model.
7- Test a new structure, go to step 4.
8- If the model is not adequate, go to
step 3 or 1.

13

Ali Karimpour Sep 2010

lecture 1

The System Identification Problem

1- Choice of Input Signals.
Filtered Gaussian White Noise.

Random Binary Noise.

Pseudo Random Binary Noise, PRBS.

Multi-Sines.

Chirp Signals or Swept Sinusoids.

Periodic Inputs.

2- Preprocessing Data.
Drifts and Detrending.

Prefiltering.

3- Selecting Model Structures.

Looking at the Data.

Getting a Feel for the Difficulties.

Examining the Difficulties.

Fine Tuning Orders and Noise Structures .

Accepting the Models .

14

Ali Karimpour Sep 2010

lecture 1

The Communities around the core

1- Statistics. ML Methods, Bootstrap method,
2- Econometrics and time series analysis.
3- Statistical learning theory.
4- Machine learning.
5- Manifold learning.
6- Chemo metrics.
7- Data Mining.
8- Artificial Neural Network.
9- Fitting Ordinary Differential equation to data.
10- System Identification.

15

Ali Karimpour Sep 2010

lecture 1

Some Open Areas in System Identification

Spend more time with neighbors.
Model Reduction and System Identification.
Issues in Identification of Non-linear Systems.
Meet Demand from Industry.
Convexification.

16

Ali Karimpour Sep 2010

lecture 1

Model Reduction
System identification is really system approximation and therefore
closely related to model reduction.
Linear systems Linear models. Divide, conquer and reunite.
Non-linear systems Linear models. Is it good for control?

Non-linear systems nonlinear reduced models. Much work remains.

17

Ali Karimpour Sep 2010

lecture 1

Linear Systems Linear Models

Divide-Conquer-Reunite
Helicopter data: 1 pulse input; 8 outputs (only 3 shown here)
State space of order 20 wanted.
x Ax bu
A is 20 20 b is 20 1
y Cx du
C is 8 20 d is 8 1

18

Ali Karimpour Sep 2010

lecture 1

Linear Systems Linear Models

Divide-Conquer-Reunite
Next fit 8 SISO models of order 12, one for each output
x i Ai xi bi u

Ai is 12 12 bi is 12 1

yi ci x d i u

ci is 1 12

d i is 1 1

i 1, 2 , ... , 8

Reunite

x Ax bu A 96 96 b 96 1
y Cx du C 8 96 d 8 1

Order reduction

19

Ali Karimpour Sep 2010

lecture 1

Linear Systems Linear Models

Divide-Conquer-Reunite
Reduce model from 96 to 20
x Ax bu A 96 96 b 96 1
y Cx du C 8 96 d 8 1

x A x bu
y C x du

A 20 20 b 20 1
C 8 20 d 8 1

20

Ali Karimpour Sep 2010

lecture 1

Convexification

21

Ali Karimpour Sep 2010