IBEHS 4A03

Section 09: Intro to Model Identification

Jake Nease
McMaster University
1
Topic Outline
▪ Black box identification of processes
– What to do when no model available
– First-order responses
– Accounting for transport delay
– Second-order (and higher) responses
– Example

▪ Limitations to using black-box models

Section09: Model_Identification 2
Empirical Model Identification
▪ Although many models exist that describe dynamical systems in biomedical and controls
engineering, sometimes estimating those parameters or the structure of the response is
difficult

▪ System Identification refers to the procedure of obtaining a transfer function of a black-

box model by applying known inputs to the system and measuring the effects
– Typically done using steps, impulses to the system
– Can be done graphically (in this course) or using data-driven modeling techniques (beyond 4A)
– If data-driven techniques used, can use more advanced input signals such a pseudo-random
binary sequences to avoid potential process drift or hysteresis error

▪ System identification is useful in engineering because either:

– We have a well-defined model of a system, but the parameters vary from patient to patient
(OGTT)
– We don’t have a well-defined system but the output we observe follows predictable dynamics and
that’s all we need to design a controller

Section09: Model_Identification 3
Empirical Model Identification
▪ So, we slap some sort of known input to the system and see what happens…

u(t) y(t)

GP??? y(t)
u(t)

▪ Based on the input sequence, we know the relationship in the ℒ ∙ domain according
to some known transfer function structure
– Easiest to establish beforehand
– Must commit to a structure before fitting parameters
– Can verify structure visually

Section09: Model_Identification 4
First-Order (Plus DT) System
▪ Objective: we would like to estimate the parameters of the following standard-form
first-order (plus dead time) system:

𝑌 𝑠 𝐾𝑒 −𝜃𝑠
=
𝑈 𝑠 𝜏𝑠 + 1

▪ We are going to apply a known step input to the system and measure the response

▪ There are lots of ways to do it

– This is just one…

Section09: Model_Identification 5
First-Order System Procedure
▪ Example: step change applied to a heat exchanger using steam (valve) as the input

NOTE… deviation
63% of new SS

Total process
gain
28% of new SS

vars
𝜽?

t63 Size of step to

t28 %
𝒖(𝒔)
%

Section09: Model_Identification 6
Workshop 9.1: FO Model Identification
𝑌′ 𝑠
▪ Determine the process transfer function 𝐺𝑃 𝑠 = using the data in the previous
𝑈′ 𝑠
plot

Section09: Model_Identification 7
Workshop 9.1 Solution
▪ Once we have data, the math is simple!

Δ 13°
𝐾= = = 𝟐. 𝟔
𝛿 5%

𝜏 = 1.5 𝑡63% − 𝑡28% = 𝟓. 𝟕

𝜃 = 𝑡63% − 𝜏 = 𝟒. 𝟏
▪ I’ll quietly remind you here that 𝜏 is roughly the time it takes to reach 63% of the
new SS

– Thus, if there IS dead time (𝜃), it is 𝑡63% − 𝜏

2.6𝑒 −4.1𝑠
𝐺𝑃 𝑠 =
5.7𝑠 + 1
Section09: Model_Identification 8
Practical Considerations
▪ Obviously, using a picture is not necessarily the best strategy
– You will likely collect data via samples, sensors, and so-on and store it in a spreadsheet

▪ There are other ways to do it

– Purely graphical approach: take the slope at the inflection point of the first-order system.
Where that intersects with 𝑌 ′ 𝑠 = 0 is the “start” of the response
– Can estimate dead time and time constant that way

▪ How large of a step in 𝑢 𝑡 ?

– Should be LARGE ENOUGH to make a measurable change in 𝑦 𝑡
– Should be SMALL ENOUGH to not break the system (cost you money, kill your patient… you
know, that sort of thing)

Section09: Model_Identification 9
Practical Considerations
▪ ALSO, as we are collecting experimental data, it is not necessarily nice or even useful

▪ Before we attempt to graphically fit a model, we must evaluate if the data is suitable
– Is the step input close enough to ideal?
– Is there noise present?
– Is the system stable?

Section09: Model_Identification 10
Workshop 9.2: Data Evaluation
▪ Evaluate whether each set of experimental data can be used to extract the
parameters of a FODT system*

*Example courtesy of Dr. Tom Marlin

Section09: Model_Identification 11
Data Evaluation Takeaways
▪ What are we looking for in experimental data?

Data Evaluation Checklist

Is the input signal nearly a perfect step?

Stable dynamic behavior

Are the assumptions of output
Overdamped dynamic behavior
behavior valid?
Monotonic, S-shaped output response

Did process begin at steady state?

Did the process achieve a new steady state?

Is the signal to noise ratio large enough (signal/noise >5)?

Does data conform to engineering fundamentals?

Was the experiment repeated, process returned to initial operation?

Section09: Model_Identification 12
Workshop 9.3: Noise and Sampling Time
▪ Suppose we have the following black-box system arranged in Simulink:

𝐺𝑝 (𝑠)

▪ We are unaware of the process dynamics, but we wish to identify a simple TF model to represent the
system
▪ Using Gaussian noise with mean, 𝜇 = 0, and standard deviation, 𝜎 = 0.2, added the response…
0.1 0.5 1 5
– Show the effect of increasing step-size on making measurable changes to 𝑦 𝑡 by using 𝑈 𝑠 = , , ,
𝑠 𝑠 𝑠 𝑠

– Show that higher (less frequent) sampling time diminishes resolution of 𝑦 𝑡

Section09: Model_Identification 13
Workshop 9.3 Solution
𝑡sample = 0.1 𝑠 𝑡sample = 1 𝑠

0.1
𝑈 𝑠 =
𝑠

0.5
𝑈 𝑠 =
𝑠

Section09: Model_Identification 14
Workshop 9.3 Solution
𝑡sample = 0.1 𝑠 𝑡sample = 1 𝑠

1
𝑈 𝑠 =
𝑠

5
𝑈 𝑠 =
𝑠

Section09: Model_Identification 15
Practical Considerations
▪ What “shape” of input 𝑢 𝑡 ?
– A step is nice because we know the shape of a first order response to a step
– Buuuut it might leave some information out
– I mean, is it really a step? (what do I mean here?)
– What about an impulse sequence?

▪ If we want to be fancier about it, we can digitize the outputs and perform a classy least-
squares regression:

𝑁
2
𝑡𝑖
min 𝜙 = ෍ 𝑦ො𝑖 − 𝑦𝑖 𝑦ො𝑖 = 𝐾 1 − exp −
𝐾,𝜏 𝜏
𝑖=1

▪ We can take ln ∙ and linearize, or we can use a nonlinear solver… Whatever. It’s all good!

Section09: Model_Identification 16
Example Data Set

Section09: Model_Identification 17
 𝑌 𝑠 𝐾
Second-Order Systems 𝑈 𝑠
= 2 2
𝜏 𝑠 + 2𝜁𝜏𝑠 + 1

▪ Second-order systems can be identified using the same properties discussed in

Section 03 (Transfer Functions)
– We can use the parameters to identify TF models
– We can ALSO use those transfer functions to estimate the unique parameters to a patient
via a parameterized model

y(t)
K
A C
1
± 5%

T
B

trise tresponse t
Section09: Model_Identification 18
 𝑌 𝑠 𝐾
Second-Order Systems 𝑈 𝑠
= 2 2
𝜏 𝑠 + 2𝜁𝜏𝑠 + 1

▪ Overshoot (𝒪 = 𝐴/𝐵)
𝑦 𝑡
– Fraction over the settling point =1
𝐾
the response peaks at y(t)
K
−𝜋𝜁 A C
𝒪 = exp
1 − 𝜁2 1
± 5%

T
▪ Decay Ratio (𝒟 = 𝐶/𝐴) B
– How quickly the oscillatory response
settles in one wave

−2𝜋𝜁
𝒟 = 𝒪 2 = exp trise tresponse t
1 − 𝜁2

Section09: Model_Identification 19
 𝑌 𝑠 𝐾
Second-Order Systems 𝑈 𝑠
= 2 2
𝜏 𝑠 + 2𝜁𝜏𝑠 + 1

▪ Period of Oscillation (T)*

– How long (in time units) one oscillation
lasts y(t)
2𝜋𝜏 K
𝑇=
1 − 𝜁2 A C
1
± 5%

▪ Response Time (tresponse) T

– Time required to settle to ± 5% of B

settling point

▪ Rise Time (trise) trise tresponse t

– Time required for system to first reach

settling point

*Note if the system is not damped (𝜁 = 0), 𝑇 is the same as the physics definition 𝑇
Section09: Model_Identification 20
Second-Order System Comments
▪ There is no unique way to estimate the parameters of a second-order system
– We pick the most convenient relations and solve for the unknowns from there

▪ If dead time is present in the system, we must account for it explicitly

– Ensure that dead time is extracted before applying second-order dynamical relations

▪ Many higher-order systems can be approximated as second-order (or even first

order)
– If we desire parameterized system responses for specific patients (remember A2?), we need
to have that system conform to a second-order system

Section09: Model_Identification 21
Workshop 9.4: Approximating as FODT
▪ State whether we can use a first order with dead time model to represent the effect
𝑇𝑀 (𝑠) *
of valve change on the measured temperature for the following process,
𝑉(𝑠)

GV 𝐹0 (𝑠) 𝑚3ൗ
𝐺𝑉 (𝑠) = = .10 𝑠൘
F0 𝑉(𝑠) % 𝑜𝑝𝑒𝑛

G1
3
𝑇1 (𝑠) −1.2 𝐾/ 𝑚 ൗ𝑠
𝐺1 (𝑠) = =
𝐹0 (𝑠) 250𝑠 + 1

T1
𝑇2 (𝑠) 1.0 𝐾/𝐾
𝐺2 (𝑠) = =
𝑇1 (𝑠) 300𝑠 + 1
G2

𝑇𝑀 𝑠 1.0 𝐾/𝐾
𝐺𝑀 (𝑠) = =
GM
𝑇2 (𝑠) 10𝑠 + 1
T

T2
TM
(Time in seconds)
*Example courtesy of Dr. Tom Marlin
Section09: Model_Identification 22
Workshop 9.4: Approximating as FODT
10
▪ If applying a step input 𝑉 𝑠 = to a system with these underlying dynamics yields
s
𝑇𝑀 (𝑠) 𝐾𝑒 −𝜃𝑠
the data seen below, determine the FODT process transfer function 𝐺𝑃 𝑠 = =
𝑉(𝑠) 𝜏𝑠+1

Section09: Model_Identification 23
Workshop 9.4: Approximating as FODT
10
▪ If applying a step input 𝑉 𝑠 = to a system with these underlying dynamics yields
s
𝑇𝑀 (𝑠) 𝐾𝑒 −𝜃𝑠
the data seen below, determine the FODT process transfer function 𝐺𝑃 𝑠 = =
𝑉(𝑠) 𝜏𝑠+1

Section09: Model_Identification 24
Higher-Order Systems?
▪ Recall, systems are limited by their largest time constant
– If that pole is dominant (produces relatively slow behaviour), we can often disregard the
effects of relatively short time constants
▪ In Workshop 9.1, we calculated 𝜏 = 5.7
𝑌 𝑠 2.6𝑒 −4.1𝑠
– With only the I/O data, how do we know the process is = versus an overdamped
𝑈 𝑠 5.7𝑠+1
higher-order process?

Section09: Model_Identification 25
Conclusions
▪ Often, we are not presented with a dynamic system of a process
– It could be such a model does not exist
– It could be that more complex models can be approximated as lesser-order
– It could be that we have a model, but need to perform tests to determine the patient’s model parameters
(OGTT, for example)

▪ Data-driven model identification has some strong advantages

– Does not require said underlying model
– Can account for many interplaying behaviours at the same time
– Can be very accurate when implemented correctly

▪ We have covered graphical methods for data-driven modeling

– Discretized systems can be solved using SSE minimization
– Yet again, optimization comes into play here

Section09: Model_Identification 26