Neural Modelling

and Control
part two
Artificial Intelligence for
Control and Identification
Dr. Wilbert G. Aguilar
Ph.D. in Automatic Control, Robotics and Computer Vision

© 2007 Dr. X. Parra & Dr. C. Angulo 0

 Outline

1. Basic Requirements

2. Classical Modelling
Neural Modelling and Control

3. Neural Modelling

4. Neural Control

1
 3. Neural Modelling

Strategies

ƒ Plant Direct Modelling

Simple structures
• Parallel
of indirect utility
Neural Modelling and Control

• Series – Parallel

ƒ Inverse Plant Direct Modelling

Inverse plant
dynamics is used
ƒ Inverse Plant Specialised Modelling to control the plant
in open loop
ƒ Backpropagation-through-time Modelling
Mimicking the
ƒ Operator Modelling control of an
expert operator

2
 3. Neural Modelling

Plant Direct Modelling: Parallel

+ Unknown ym(t)
um(t) +
Plant
Neural Modelling and Control

Neural Network y(k)

–+
Plant Model εy(k) = ym(k) – y(k)
Jy = E(εy2(k))
BP

ƒ Identical to classical modelling by regression analysis

ƒ Simple but rarely used because
• ANN parameters convergence is not guaranteed
• It doesn’t ensure that training process error tends to zero
3
 3. Neural Modelling

Plant Direct Modelling: Series - Parallel

+ Unknown ym(t)
um(t) +
Plant
Neural Modelling and Control

Neural Network y(k)

–+
Plant Model εy(k) = ym(k) – y(k)
Jy = E(εy2(k))
BP

ƒ Dynamical identification in series with the output signal and in

parallel with the input signal

ƒ Both, ANN parameters convergence and zero error in steady

state are guaranteed
4
 3. Neural Modelling

Plant Direct Modelling: Application

ƒ Off-line strategy in control design.

ƒ Estimating the plant output into a larger

Neural Modelling and Control

closed loop control.

ƒ Predicting plant features when the real

output is unavailable (for instance, due to
time delays on sensors).

ƒ Fault diagnosis.

5
 3. Neural Modelling

Plant Direct Modelling: Drawback

Main Problem

“A poor or scarcely representative training set

Neural Modelling and Control

will generate a biased model towards those

overtrained input areas”

Pay attention on choosing

equidistant training data on
relevant areas of the input space

6
 3. Neural Modelling

Inverse Plant Direct Modelling

+ Unknown ym(t)
um(t) +
Plant
Neural Modelling and Control

+– u(k) Neural Network

εu(k) = um(k) – u(k) Inverse Model
Ju = E(εu2(k))
BP

ƒ Goal: Design a controller such that the whole system

[plant (unknown) + controller (inverse plant model)]
has unitary transfer function

7
 3. Neural Modelling

Inverse Plant Direct Modelling

Warning

“Modelling errors cause that the whole system

Neural Modelling and Control

deviates from unitary transfer function”

Good performance for a wide range of nonlinear

plants is obtained when the inverse plant model is
used as a feed-forward pre-compensator jointly with a
linear standard controller (PID) in the feedback loop

8
 3. Neural Modelling

Inverse Plant Direct Modelling: Drawback

Main Problem

“Learning is not directly oriented

Neural Modelling and Control

to the controlled signal ym(t)”

ƒ It is quite difficult to determine a suitable range for

the control variable (u) such that the controlled
variable (y) is conveniently covered.
ƒ MSE is minimized in the control space rather than
in the output space.

9
 3. Neural Modelling

Inverse Plant Direct Modelling: Drawback

ƒ MSE is minimized in the control space rather than in the output

space, so
∂y(t) ⎛ ⎛ ∂y(t) ⎞
2
⎞
ε
y (t) ≈ ε⋅ u (t) ⇒ J y ≈ E ⎜ ⎜ ε ⋅ u (t) ⎟
Neural Modelling and Control

2
⎟
⎜ ⎝ ∂u(t) ⎠ ⎟
∂u(t) ⎝ ⎠
ƒ For a nonlinear plant, minimizing MSEu does not imply necessarily
minimizing MSEy

ƒ However, when output errors are not correlated with the Jacobian of
the plant, then
⎛ ⎛ ∂y(t) ⎞2 ⎞
Jy ≈ E ⎜ ⎜ ⎟⋅J
⎜ ⎝ ∂u(t) ⎟⎠ ⎟ u
⎝ ⎠
and the Jacobian effect can be incorporated on the learning rate, η

10
 3. Neural Modelling

Inverse Plant Specialised Modelling

Neural Network + Unknown ym(t)

yd(t) Inverse Model
ZOH +
Plant
Neural Modelling and Control

u(k) Neural Network y(k)

BP
Ju = E(εu2(k))
Plant Model

εy(k) = [yd(k) – y(k)] or [yd(k) – ym(k)]

+ –

ƒ Learning is directly oriented to the controlled signal ym(t)

ƒ Training is performed in two phases

11
 3. Neural Modelling

Inverse Plant Specialised Modelling

Training phases
i. A direct plant model is built covering all the operations range
Neural Modelling and Control

ii. Inverse plant model is built in two steps:

a. Start using the measured output signal ym(t), except for

very noisy cases

b. Continue using the plant modelled output y(t) (plant

model must cover whole range of operation)

12
 3. Neural Modelling

Inverse Plant Specialised Modelling

Advantage

“For on-line applications, model output error drives inverse model

Neural Modelling and Control

to unexplored zones, where its training will be performed, whereas

direct model was trained on different control areas”

Drawback

“Backpropagation algorithm cannot be directly used because

driving error [yd(k) – y(k)] or [yd(k) – ym(k)] is not directly related
with the neural network output u(k), so it must be supposed that…”
⎛ ⎛ ∂y(t) ⎞2 ⎞
Ju ≈ J y E⎜⎜ ⎟
⎜ ⎝ ∂u(t) ⎟⎠ ⎟
⎝ ⎠

13
 3. Neural Modelling

Backpropagation-through-time Modelling
z-1

Neural Network + Unknown ym(t)

yd(t) Inverse Model
ZOH +
Plant
Neural Modelling and Control

u(k) Neural Network y(k)

BP
Plant Model

εu(k)
BP

εy(k)
+ –
ƒ An emulation of εu(k) is obtained,
backpropagating εy(k) towards the input layer of the ANN plant model
14
 3. Neural Modelling

Backpropagation-through-time Modelling

Remarks

ƒ This method assumes that input u(t) has a linear effect on the
Neural Modelling and Control

unknown process. It is a hard assumption, but it makes able

the control for many cases.

ƒ The process model is trained before the inverse model (z-1).

ƒ Although error εy(k) is backpropagated through the neural

network plant model, this model is not adapted during the
training of the neural network inverse plant model.

ƒ A possible variant is to use a reference model to insert into

the BTT model a transient desired output instead of a fixed
reference.

15
 3. Neural Modelling

Operator Modelling (in closed loop)

yd(t) – Expert um(t) + Unknown ym(t)

+ +
Operator Plant
Neural Modelling and Control

Neural Network u(k)

–+
Expert Model

BP

ƒ Controller is learned from an expert: learning algorithm is

executed in parallel to a plant operator and his/her operations
are taken as the neural network training targets.
16