Cccshort 6

1 The Problem 2
540
520
500
Approaches to Identification of
480
0 20 40 60 80 100
t (s)
560
540
520
Nonlinear systems 500
480
0 20 40 60 80 100
t (s)
Brain Activity (fMRI):
OUTPUT #1
25
20
15
10
5
0 100 200 300 400 500 600
INPUT #1
20
Lennart Ljung 15
10
0 100 200 300 400 500 600
Pulp Buffer Vessel:

OUTPUT #1
4
Division of Automatic Control -2
-4
0 100 200 300 400 500 600
INPUT #1
1.5
0.5
Linköping University 0
-0.5
-1
Forest Crane: 0 100 200 300 400 500 600
Sweden
Nonlinear Identification CCC Plenary AUTOMATIC CONTROL Nonlinear Identification CCC Plenary AUTOMATIC CONTROL
REGLERTEKNIK REGLERTEKNIK
Lennart Ljung July 29, 2010 Lennart Ljung July 29, 2010
LINKÖPINGS UNIVERSITET LINKÖPINGS UNIVERSITET
This Presentation ... 3 Dynamic Systems 4
... aims at A dynamic system has an output

response y that depends on (all)
■ an explanation of the essence of the problem of non-linear identification previous values of an input sig-
■ a color-coded overview of typical parametric approaches nal u. It is also typically affected
■ some information of on-going research on a new, non-parametric by a disturbance signal v.
approach So the output at time t can be
written as
y(t) = g(ut , v t )
where superscript denotes the

signal’s values from the remote
past up to the indicated time.
The input signal u is known
(measured), while the distur-
bance v is unmeasured.
More Formalized Questions 5 The Predictor 6
Think discrete time data sequences: The predictor function ŷ(t|t − 1) = f˜(Z t−1 ) = f (ϕ(Z t−1 )) is what we try to
Z t = [u(1), u(2), ..., u(t), y(1), y(2), ..., y(t)] estimate from data.
We need to get hold of a “sim- and/or a prediction func-
ulation function” y(t) = g(ut ) tion ŷ(t|t − 1) = f˜(Z t−1 ) Two basic cases for f : (special case ϕ(t) = [u(t − 1), u(t − 2)])
5
x 10
3 4
2.5 3
2 2
*
1.5 1
******* **
******** ******
1 * ********* 0
y
y
************* ************
0.5
** ********* −1
* ********
** *** ***********
* * **** ** ****
0 −2 *** **
******** **
−0.5 −3
*******
*
−1 −4
6 6
5 5
5 5
4 4
4 4
3 3
3 3
2 2 2 2
1 1 1 1
0 0 0 0
u(t−2) u(t−2)
u(t−1) u(t−1)
Linear: ŷ(t) = a1 u(t − 1) + a2 u(t − 2) Nonlinear: ŷ(t) = f (u(t − 1), u(t − 2))
The observations Z t are points in this space.

Note that f ⇒ g!
A Quick Classification of Non-Linear Models 7 White to Off-White: Physical Modeling 8
Perform physical modeling (e.g. in M ODELICA) and denote (unknown)

physical parameters by θ. Collect the model equations as
5
x 10
2
ẋ(t) = f (x(t), u(t), θ)
1 **
******
0
y
***********
−1
* ******** y(t) = h(x(t), u(t), θ)
***********
*** *****
−2 ** *
* *
−3 ******
−4
6 (or in DAE, Differential Algebraic Equations, form.) For each parameter θ this
5
5
4
4
3
3
2 2
defines a simulated (predicted) output ŷ(t|θ) which is the parameterized
1 1
0 0
u(t−2)
How to describe the surface u(t−1) ? function

■ Parametric: fˆ(ϕ) = f (ϕ, θ̂) θ̂ = arg min y(t) − f (ϕ(t), θ)2 ŷ(t|θ) = f˜(Z t−1 , θ)
■ Nonparametric: Form the surface by smoothing the observations y(t)
in somewhat implicit form. θ is then found by optimizing the fit to
observations.
Parametric Models: A Palette from White to Black: The approach is conceptually simple, but could be very demanding in
practice.
White – Off-white – Smoke-grey – Steel-grey –Slate-grey – Black
The Palette 9 Smoke-Grey: Semi-physical Models 10
■ White: Known model Apply non-linear transformations to the measured data, so that the
■ Off-white: Careful Physical Modeling transformed data stand a better chance to describe the system in a linear
relationship.
■ Smoke-grey
■ Steel-grey “Rules: Only high-school physics and max 10 minutes”
■ Slate-grey
■
Toy Example: Immersion heater: Input: voltage to the heater. Output:
Black
temperature of the fluid. . . ..
. . .. Square the voltage! Sense morale: No excuse for not thinking over the
basic physical facts!
Another example: . . .
Buffer Vessel Dynamics 11 Model Based on Raw Data 12
OUTPUT #1
25 80
20 60 Measured Output and Simulated Model Output

6
Measured Output
15 40 mraw Fit: 21.11%
10 20 4
5 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
2
INPUT #1
200
0
20 150
y1
100
15 −2
50
10
0 −4
0 100 200 300 400 500 600 0 100 200 300 400 500 600
−6
κ-number of outflow, flow

−8
volume 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
4
κ-number of inflow, x 10
Dashed line: κ-number after the vessel, actual measurements.

Solid line: Simulated κ-number using the input only and a process model
0.818 −480s
estimated using the first 200 data points. G(s) = 1+676s e
Now it’s time to 13 Re-sample Data 14
Think: .... z = [y,u]; pf = flow./level;

t = 1:length(z)
No mixing (“Plug flow”): The vessel is then just a pure time delay for the pulp newt = interp1([cumsum(pf),t],[pf(1):sum(pf)]’);
flow: Delay time: Vessel Volume/Pulp Flow (dimension time.) newz = interp1([t,z], newt);
κ−number of Inflow
25
Perfect mixing in tank: A text-book first order system with gain=1 and time 20
constant = Volume/Flow 15
10
So if Volume and Flow are changing, we have a time-varying system (or 5

0 10 20 30 40 50 60 70 80
non-linear!)
κ−number of Outflow
22
20
The natural time variable is really Volume/Flow, (which we have measured). 18
16
Let us re-sample the observed data according to this natural time variable. 14
12
10
8
0 10 20 30 40 50 60 70 80
Semi-physical Model 15 The Palette 16
■ White: Known model

0.8116
Measured Output and Simulated Model Output
G(s) = e−369.58s 6
Measured Output
■ Off-white: Careful Physical Modeling
1 + 110.28s mraw Fit: 21.11%
2 ■
Measured Output and Simulated Model Output Smoke-grey. Semi-physical modeling (Could be used more!)
6 0
Measured Output
y1
mves Fit: 60.39% −2
■ Steel-grey
−4
4
−6
■ Slate-grey
−8
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
4
x 10
2 ■ Black
Recall Linear model.
y1
−2 The semi-physical model gives
a sufficiently good description
−4
of the buffer, to allow proper
−6
time-marking of the pulp before
and after.
−8
0 5000 10000 15000
Steel-Grey: Composite Local Models 17 Tank Example, ctd 18
Non-linear systems are often handled by linearization around a working point.

The idea behind Composite Local (Local Linear) Models is to deal with the
nonlinearities by selecting or averaging over relevant linearized models. Sampled data model around level h∗ (Sampling time Ts ):
Example: Tank with inflow u and free outflow y and level h: (Bernoulli’s) y(t) = γ(h∗ ) + α(h∗ )y(t − Ts ) + β(h∗ )u(t − Ts ) = θ T (h∗ )ϕ(t)
equations:
An ARX-model with level-dependent parameters. Now compute linearized
√
ḣ = − h + u model for d different levels, h1 , h2 , . . . , hd . Total model: select or average over
√ these local models
y= h
d

√ ŷ(t) = wk (h, hk )θ T (hk )ϕ(t)
Linearize around level h∗ with corresponding flows u∗ = y ∗ = h∗ :
k=1
1
ḣ = − √ (h − h∗ ) + (u − u∗ ) Choices of weights wk : . . ..
2 h∗
1
y = y ∗ + √ (h − h∗ )
2 h∗
Data and Linear Model 19 Local Linear Models 20
Measured data: Linear Model (d = 1) Two levels (models) (d=2) Five levels (models) (d = 5)
Inflöde u 1.4 1.4
1.4 1.4
1.2
1.2 1.2
1
1.2
0.8
0.6 1 1
1
0.4
0.2
0.8 0.8
0 0.8
0 10 20 30 40 50 60 70 80
Nivå h 0.6 0.6

1.5 0.6
0.4 0.4
1 0.4
0.2 0.2
0.5 0.2
0 0
0 0 0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
0 10 20 30 40 50 60 70 80 0 10 20 30 40 50 60 70 80
Thick line: Model. Thin: Measured.
Composite Local Models: General Comments 21 The Palette 22
Let the measured working point variable (tank level in example) be denoted ■ White: Known model
by ρ(t) (sometimes called regime variable). If the regime variable is ■ Off-white: Physical Modeling
partitioned into d values ρk , and model output according to value ρk is ŷ (k) (t) ■ Smoke-grey: Semi-physical modeling
the predicted output will be
■ Steel-grey: Composite (local) models (Most common NL model in
d
industry(?))
ŷ(t) = wk (ρ(t), ρk )ŷ (k) (t) ■ Slate-grey
k=1
■ Black
If the prediction ŷ (k) (t) corresponding to ρk is linear in the parameters,
ŷ (k) (t) = ϕT (t)θ (k) , and the weights w are fixed, the whole model will be a
linear regression.
Important connections to active research areas
■ LPV (Linear Parameter-Varying) Models
■ Hybrid Models (≈ w(·, ·) is estimated too.)
Slate Grey: Block-oriented Models 23 Common Models 24
Building Blocks:
Wiener
Linear Dynamic System

G(s)
Hammerstein
Hammerstein-
Wiener
Nonlinear static function
f (u)
Other Combinations 25 Example: Hydraulic Crane Data 26
Active Research Field:

These are data from a forest harvest machine:
OUTPUT #1
4
-2
-4
0 100 200 300 400 500 600
INPUT #1
1.5
0.5
With the linear blocks parameterized as a linear dynamic system and the
0
static blocks parameterized as a function (“curve”), this gives a -0.5
-1
parameterization of the output as 0 100 200 300 400 500 600
Input: Hydraulic Pressure.

t−1
ŷ(t|θ) = g(Z , θ) Output: Tip Position
and the general approach of model fitting can be applied.

However, in this contexts many algorithmic variants have been suggested.
Linear Model 27 Hammerstein Model of the Hydraulic Crane 28
Black: Measured Output Hammerstein model: Fit 71.61 %

4
Blue: Model Simulated Output
3 The Hammerstein
Linear model: Fit 41.71 %
4 Model gives a good
2
1
fit. The extra flexi-
3
bility offered by the
0 input nonlinearity
2
−1 is quite useful,
1
−2 (even though no
0
−3
direct physical
explanation is
−1 −4
obvious.)
−5
−2
−6
0 200 400 600 800 1000 1200
−3
−4
0 200 400 600 800 1000 1200
The Palette 29 Black: Basis-Function Expansion 30
■ White: Known model ŷ(t|θ) = f˜(Z t−1 , θ) = f (ϕ(t), θ)

■ Off-white: Physical Modeling
ϕ(t) = ϕ(Z t−1 ) "state" of fixed dimension
■ Smoke-grey: Semi-physical modeling
d

■ Steel-grey: Composite (local) models f (ϕ, θ) = αk κ(βk (ϕ − γk )), θ = {αk , βk , γk } κ : unit function
■ Slate-grey: Block-oriented models k=1
■ Black
Intuitive picture: Think of a scalar ϕ and let κ(z) be a unit pulse for
0 ≤ z ≤ 1. Then κ(β(ϕ − γ)) is a pulse of width 1/β starting in ϕ = γ. The
sum above is then a piecewise constant function, capable of approximation
"any" function arbitrary well for large enough d.
■ The whole ANN (Artificial Neural Network), neuro-fuzzy, LS-SVM (Least

Squares Support Vector Machines), etc business
Non-Parametric Methods 31 Semi-supervised Regression: The Problem 32
5
x 10
Given a standard regression problem:
4
1 **
****** y(t) = f (ϕ(t)) + noise
0
y
************
−1
* ********
***********
*** *****
−2
*** **
−3 ******
−4
6 f unknown, y(t) and ϕ(t) observed for t = 1, . . . , NL . Find f !
5
5
4
4
3 3
2 2
1 1
0 0
u(t−2)
u(t−1)
Form the surface by smoothing over the observation points in the space!
Or rather, for any given ϕ∗ find a good value of f (ϕ∗ ) (“Model on Demand”,
■ Even Blacker!
“Just in time model”)
■ Huge literature – Mostly in the statistical community and now also in
machine learning
■ Only one aspect will be discussed here: Semi-supervised Regression Extra feature: We have several measurements of
ϕ(t), t = NL + 1, . . . , NL + NU without corresponding values of y(t) (NU
“unlabeled observations”)
Information in Unlabeled Regressors? 33 (Un-/Semi-)Supervised 34
Yes, if the regressors live in a confined region, like a manifold: ■ Supervised: All Regressors labeled (y(t), ϕ(t)) [Standard Regression
problems]
■ Unsupervised: No labels known. [Clustering, Classification]
■ Semi-supervised: Some labels known {y(t), ϕ(t), t = 1, . . . , NL }.
Some additional regressors known without labels
{ϕ(t), t = NL + 1, . . . , NL + NU }
■ Estimation problem: still to “predict” y ∗ = f (ϕ∗ ) for any given ϕ∗
• ”predict”: .....
The Suggested Method: WDMR (With H. Ohlsson) 35 WDMR, cont’d 36
WDMR: Weight Determination by Manifold Regression Now, weigh together the two sources of information
(Manifold) Smoothness Assumption: f (ϕ1 ) and f (ϕ2 ) close if ϕ1 and ϕ2 are
close (in a relevant metric). +NU
NL +NU
NL NL

λ (fˆi − Ki,j fˆj )2 + (1 − λ) (y(t) − fˆt )2 (∗)
Problem: associate ϕ(i) with good values fˆt = f (ϕ(t)) for all regressors, both i=1 j=1 t=1
labeled and unlabeled. Ki,j = K(ϕ(i), ϕ(j))
Take care of two kinds of information:
■ f (ϕ(t)) ≈ y(t) for the measured labels Pick a (“regularization parameter”) λ that balances the fit to measured labels
■ f (ϕ(t)) ≈ f (ϕ(j)) if ϕ(t) ≈ ϕ(j). and the smoothness prior. Minimize w.r.t. fˆt , t = 1, . . . , NL + NU .
Formalize the second information using a kernel K(·, ·): That gives the estimated function value fˆ(ϕ) for any regressor ϕ you include
among the unlabeled regressors.
+NU
NL
Note that (∗) is quadratic in fˆt , so the solution is easy to obtain.
fˆt ≈ Kt,j fˆj ; Kt,j = K(ϕ(t), ϕ(j))
Choice of kernel K(·, ·): Many possibilities ...
t=1
Example 1: fMRI Signals 37 fMRI Signals, cont’d 38
80 (labeled) samples were collected as estimation data. The model is a

The patient in the magnet cam- mapping from the ϕ-space R128 to the scalar space of viewing angle [±π].
era is moving his eye focus left For the mapping in this large space. 80 measurements may seem to be very
- right - up - down. 128 voxels few. But the brain activity in the 128-dimensional space is really triggered by
in the visual cortex are moni- a one-dimensional stimulus - the scalar eye movement. It can thus be argued
tored by fMRI, giving a vector that the regressors in R128 are really confined to a one-dimensional manifold.
ϕ(t) ∈ R128 sampled every two To test the method, also 40 validation data were collected. (i.e.
seconds. The output y(t) is the NL = 80, NU = 40)
viewing angle 0, π, π/2, −π/2.
To the right we show the pre-
dicted y-values ([−π/2 π]),
(thick line) for these unlabeled
validation measurements to-
gether with the corresponding
true angles (thin line).
Conclusions 39
■ The world of non-linear identification is rich and complex.

■ Parametric methods may be color-coded in several shades of grey
■ Non-parametric methods are gaining importance with inspiration from
statistics and “machine learning”. They certainly have relevance for
system identification
■ A “semi-supervised” method WDMR was suggested for non-linear,
non-parametric regression, which shows promising results for several
examples of different characters (but needs more understanding
regarding the potential for system identification)
Nonlinear Identification CCC Plenary AUTOMATIC CONTROL

REGLERTEKNIK
Lennart Ljung July 29, 2010
LINKÖPINGS UNIVERSITET

Cccshort 6

Enviado por

Dados do documento

Descrição original:

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Cccshort 6

Enviado por

Direitos autorais:

Formatos disponíveis

1 The Problem 2

0 100 200 300 400 500 600

Pulp Buffer Vessel:

Division of Automatic Control -2

This Presentation ... 3 Dynamic Systems 4

... aims at A dynamic system has an output

where superscript denotes the

More Formalized Questions 5 The Predictor 6

The observations Z t are points in this space.

Perform physical modeling (e.g. in M ODELICA) and denote (unknown)

The Palette 9 Smoke-Grey: Semi-physical Models 10

Buffer Vessel Dynamics 11 Model Based on Raw Data 12

20 60 Measured Output and Simulated Model Output

κ-number of outflow, flow

Dashed line: κ-number after the vessel, actual measurements.

Think: .... z = [y,u]; pf = flow./level;

So if Volume and Flow are changing, we have a time-varying system (or 5

The natural time variable is really Volume/Flow, (which we have measured). 18

Semi-physical Model 15 The Palette 16

■ White: Known model

Steel-Grey: Composite Local Models 17 Tank Example, ctd 18

Non-linear systems are often handled by linearization around a working point.

Nivå h 0.6 0.6

Thick line: Model. Thin: Measured.

Composite Local Models: General Comments 21 The Palette 22

Slate Grey: Block-oriented Models 23 Common Models 24

Linear Dynamic System

Active Research Field:

Input: Hydraulic Pressure.

and the general approach of model fitting can be applied.

Linear Model 27 Hammerstein Model of the Hydraulic Crane 28

Black: Measured Output Hammerstein model: Fit 71.61 %

The Palette 29 Black: Basis-Function Expansion 30

■ White: Known model ŷ(t|θ) = f˜(Z t−1 , θ) = f (ϕ(t), θ)

■ The whole ANN (Artificial Neural Network), neuro-fuzzy, LS-SVM (Least

Information in Unlabeled Regressors? 33 (Un-/Semi-)Supervised 34

The Suggested Method: WDMR (With H. Ohlsson) 35 WDMR, cont’d 36

80 (labeled) samples were collected as estimation data. The model is a

■ The world of non-linear identification is rich and complex.

Nonlinear Identification CCC Plenary AUTOMATIC CONTROL

Você também pode gostar