System Identification

Lecture 1: Introduction

Roy Smith

2019-9-12 1.1

Identification objectives

§ Simple approximation to the system.

§ Model for prediction purposes.
§ Clearer understanding of system behaviour.
§ Model for use in control design.

2019-9-12 1.2
Fundamental issues

§ Purpose of the model.

§ Open- or closed-loop experiments.
§ Linearity of the system. Linear model assumptions.
§ Choice of sample period, T .
§ Finite data records.
§ Constraints on the input excitation.
§ Noise on the measurements.
§ Experiment time constraints.

2019-9-12 1.3

Fundamental concepts

ypkq upkq
G

An Identification Procedure

1. Choose an input, upkq, k “ 1, . . .

2. Apply upkq to the plant, G.
3. Measure ypkq, k “ 1, . . . .
4. Use tupkq, ypkqu, k “ 1, . . . to get a model for G.

Basic idea:

y “ Gu, so Ĝ « y{u.

2019-9-12 1.4
Models

First principles (parametric)

Derived from physical principles.
1{m
f “ m:
x ùñ Gpsq “ , where y “ x, and u “ f.
s2

“Black-box” models
Response estimated from experimental data (measurements of f ptq and xptq).

“Grey-box” models
A combination of first principles and experimentally derived parameters.
θ
Gpsq “ , where θ is to be estimated.
s2

2019-9-12 1.5

Frequency-domain and time-domain methods

Frequency domain:
§ Estimate U pejω q and Y pejω q from upkq and ypkq.
§ Estimate Ĝpejω q from U pejω q and Y pejω q.
§ Averaging and windowing can give very good control over the treatment
of noise corruption.
§ Usually requires more data than time-domain methods.

Time-domain:
§ Use Z-transform time relationships directly: ypk ` 1q “ zypkq.
§ Estimate coefficients of Z-domain polynomial from data.
§ Often requires fewer data than frequency domain to give an estimate.
§ It is more difficult to handle the effects of noise.

2019-9-12 1.6
Model purpose

Open-loop prediction Closed-loop design

y u r
ŷ Ĝ C `
u
Ĝ ´

ŷ “ Ĝu. y ĜC
“
r 1 ` ĜC

GC
Criteria: stable
Criterion: }y ´ ŷ} I ` GC
› ›
› ›
› GC ĜC ›
› ´ ›
› 1 ` GC 1 ` ĜC ›

2019-9-12 1.7

Model purpose

Model comparison: frequency domain

Magnitude
100

G2 pjωq Gpjωq
0.1 1 10 Frequency, ω
1 (rad/sec)

G1 pjωq

0.01

0.1 1 10 Frequency, ω
0 (rad/sec)
G2 pjωq
´90

´180
Gpjωq G1 pjωq
´270
Phase (deg.)

2019-9-12 1.8
Model purpose

Step responses
Model outputs
50
żt
yptq “ gpτ qupt ´ τ qdτ
40 0

30 żt
y1 ptq “ g1 pτ qupt ´ τ qdτ
0
20 żt
y2 ptq “ g2 pτ qupt ´ τ qdτ
0
10

0 time (sec.)
5 10 15 20 25 30

2019-9-12 1.9

Model purpose

Closed-loop responses
Magnitude Controller:
10
0.01p10s ` 1q
Cpsq “
1 10
p0.1s ` 1q
Frequency, ω
1 (rad/sec)
|G2 pjωqCpjωq|
|1 ` G2 pjωqCpjωq|
0.1 |G1 pjωqCpjωq|
|1 ` G1 pjωqCpjωq|

0.01 |GpjωqCpjωq|
|1 ` GpjωqCpjωq|

0.001

2019-9-12 1.10
Model purpose

Closed-loop responses
Magnitude Controller:
10
0.01p10s ` 1q
Cpsq “
1 10
p0.1s ` 1q
Frequency, ω
1 (rad/sec)
|G2 pjωqCpjωq|
|1 ` G2 pjωqCpjωq|
0.1 |G1 pjωqCpjωq| Closed-loop stability:
|1 ` G1 pjωqCpjωq| G1 stable
0.01 |GpjωqCpjωq|
G2 unstable
|1 ` GpjωqCpjωq| G unstable

0.001

2019-9-12 1.10

Model purpose

Relative errors

Magnitude
10
|Gpjωq ´ G1 pjωq|
|Gpjωq|
Frequency, ω
1 1 10 (rad/sec)

0.1

|Gpjωq ´ G2 pjωq|
0.01 |Gpjωq|

0.001

2019-9-12 1.11
Open-loop identification

Discrete-time model:
epkq

Hpejω q

vpkq
ypkq upkq
` Gpejω q

Main assumptions
§ G is linear.
§ epkq is zero-mean white noise (with known variance).
§ H may be known approximately.

2019-9-12 1.12

Closed-loop identification

epkq

Hpzq
r2 pkq
vpkq
ypkq upkq r1 pkq
+ Gpzq ` Cpzq `
´

Key features
§ Closed-loop operation can set the operating point of the plant.
§ For many plants closed-loop operation is required.
§ For unstable plants, closed-loop experiments are essential.
§ A well designed controller masks variations in the plant.
§ We can now measure ypkq, upkq, r1 pkq and r2 pkq.
§ All of these signals are now correlated.

2019-9-12 1.13
46 kWh/m2 pro Jahr
Example: energy flows in buildings
auch 83 kWh/m2 pro Jahr

winnung,
ng

urm

Inputs: Air supply temperatures

In-floor water system temperatures
Blind settings
Solar radiation
Outside temperature

Outputs: Room temperatures

2019-9-12 1.14

Example: energy flows in buildings

Room temperatures (controlled operation)

Temperature (˝ C)
28
Upper comfort bound

26
Maximum
Average
Minimum
24

22
Lower comfort bound

20 Date
1 May 8 May 15 May 22 May 29 May 4 Jun 11 Jun 18 Jun

2019-9-12 1.15
Step (or doublet) excitation

Step changes in TABS heating setpoint

Excitation plan: 16 ˝ C for 5 days

30 ˝ C for 5 days

Temperature (˝ C)

28

24

20

16

Date
24 Dec. 28 Dec. 01 Jan. 05 Jan.

2019-9-12 1.16

Example: energy flows in buildings

Things to consider:

§ The building is occupied and (usually) controlled.

§ The dynamics can be (very) slow.
§ It takes a year to see a reasonably complete range of weather conditions.

2019-9-12 1.17
Example: vibration in flexible strucures
» fi
y1
–y2 fl
y3

Inputs: Forces in structural members. (ui )

Shaker disturbance (w)
w

Outputs: Accelerations in 3 dimensions (yi )

u1
u2
u3

2019-9-12 1.18

Example: vibration in flexible structures

Magnitude
100

10

Frequency, ω
1 (rad/sec)
40 50 60 70 80 90

0.1

0.01

0.001

2019-9-12 1.19
Example: vibration in flexible structures
Magnitude
1000

100

10

Frequency, ω
1 (rad/sec)
50 100 200 500

0.1

0.01

0.001

2019-9-12 1.20

Example: vibration in flexible structures

Things to consider:

§ Extremely fine frequency resolution is needed to resolve the modes.

§ Long data records are required.
§ There is noise and drift on the accelerometer measurements.
§ Experiments are very sensitive to environmental disturbances.

2019-9-12 1.21
Example: process control
hot fh
actuator

cold fc
actuator

height:
h1

temperature: t1

output flow: f1

Inputs: Hot and cold water flows (fh and fc )

Outputs: Tank height and temperature (h1 and t1 )

2019-9-12 1.22

Example: process control

Height response: fh ` fc to h1

10 Magnitude
0.001 0.01 0.1 1 Frequency, ω
1 (rad/sec)
0.1

0.01

0.001

0.0001 model: P

0.001 0.01 0.1 1 Frequency, ω

0 (rad/sec)
´90

´180

´270

´360

´450 model: P
Phase (deg.)
2019-9-12 1.23
Example: process control

Temperature response: fh ´ fc to t1

10 Magnitude
0.001 0.01 0.1 1 Frequency, ω
1 (rad/sec)
0.1 h1 “ 0.15

0.01

0.001 h1 “ 0.75

0.0001

0.001 0.01 0.1 1 Frequency, ω

0 (rad/sec)
´90

´180

´270
h1 “ 0.75
´360
h1 “ 0.15
´450 Phase (deg.)
2019-9-12 1.24

Example: process control

Things to consider:

§ Temperature and pressure (height) measurements are noisy.

§ A wide range of frequencies were identified (multiple experiments).
§ Nonlinear mixing dynamics are significant in the temperature response.
§ The temperature response is a strong function of the height.
§ Closed-loop control was used for one variable (h1 or t1 ), while the other
was excited in open-loop.

2019-9-12 1.25
Sampled-data configuration

A typical practical configuration

eptq

Hpsq

vptq
ypkq yptq uptq upkq
` Gpsq ZOH
T

ZOH block: Digital-to-analog converter (DAC)

sampling (T ) block: Analog-to-digital convertor (ADC)

The selection of the sampling time, T , will determine the maximum frequency
that we can model with a discrete-time system.

2019-9-12 1.26

Sampled-data configuration

An equivalent model:

eptq

Hpsq

vptq
ypkq yptq uptq upkq
` Gpsq ZOH
T

epkq

Hpejω q

vpkq
ypkq upkq
` Gpejω q

2019-9-12 1.27
Sampling operation
ˇ xpkq xptq
ˇ
ypkq “ yptqˇˇ T
t“kT,k“0,1,2,...
T is the sampling period.

4
ypkT q
3

2 yptq
Continuous signal:
1
Time
0 (seconds)
1T 2T 3T 4T 5T 6T
´1

4
ypkq
3

2
Discrete sequence:
1
Sample
0
1 2 3 4 5 6 index
´1

2019-9-12 1.28

Discrete frequencies: Fourier series of periodic signals

Periodic signal: xpkq (period = M samples, assume M is even).

The Fourier series is:

ÿ
M ´1
2πm 2π
Xpe jωm
q “ xpkqe´jωm k , where ωm “ “ mω0 , ω0 “
k“0
M M
m “ 0, . . . , M ´ 1.

This is also periodic (with period = M ).

Non-negative frequencies are: m “ 0 to m “ M {2,

2π 4π 2πpM {2 ´ 1q
corresponding to: ωm “ 0, , ,..., , π.
M M M

2019-9-12 1.29
Sampled periodic signals

Sampling operation:
xpkq xptq
Periodic signal with period = τp , T

xptq “ xpt ` τp q, t P p´8, 8q.

Now sample xptq with sampling time: T

To get M samples per period (i.e. M T “ τp ),

τp
T “ .
M
This gives frequencies,
ˆ ˙ ˆ ˙ ˆ ˙
2π 2π 2π M 2π
0, , 2 ,3 ,¨¨¨ ,
τp τ p τ p 2 τp
loomoon loooooooooooooooooooooomoooooooooooooooooooooon
Fundamental
Harmonics
frequency

2019-9-12 1.30

Sampled periodic signals

Lowest frequency:
1.0 ypkT q
2π
0.5 ωu “ ω1 “
yptq Time (sec.) MT
0
T 2T 3T 4T 5T 6T 7T 8T 9T 10T (apart from zero)
´0.5

´1.0 M “ 10

ypkT q
1.0

0.5 yptq
Time (sec.)
0 Highest frequency:
T 2T 3T 4T 5T 6T 7T 8T 9T 10T
´0.5 π
ωu “ ωM {2 “
´1.0 T
(Nyquist frequency)

There are only M {2 ` 1 distinct frequencies (counting zero) that can be

represented in a sampled signal.

2019-9-12 1.31
Time and frequency domain sampling relationships

The sampling time, T , determines the Nyquist frequency π{T .

The experiment length, KT , determines the frequency resolution, 2π{KT .

0 T 2T pK´1qT
Time
(sec.)

inverse
DFT
DFT

Frequency
(radians/sec.)
0 2π π 2π
KT T T
Negative
frequencies

2019-9-12 1.32