URUAPAN HIGHER TECHNOLOGICAL INSTITUTE

MECHATRONICS ENGINEERING
9th Semester, Group “A”

PROCESS CONTROL

“REFERENCE MODEL METHOD”

TEACHER: ENGINEER ROGELIO ESQUEDA GARCIA

MEMBERS:
Alves Flores Isidro
Coria Rivera Jesus Alejandro Leon
Martinez Alejandra Santoyo
Rivas Luis
Alberto Virrueta Martinez Enrique
URUAPAN, MICHOACÁN, SEPTEMBER 5, 2016.
Model reference control is a generalization of step predictive
controllers. What is attempted to be done in this case is to consider in
the design the dynamics of the reference, that is, the model to follow.
There are two views of the problem: one deterministic and the other
stochastic depending on the plant model.

■ Control with Reference Model

> Control with Adaptive Reference Model
> Control with Stochastic Reference Model

Reference model method

Of the conceptual systems identification methods, one of the

simplest, and at the same time the most flexible, is the reference
model or also called learning models. The basic idea of his approach
is represented in the figure.
 y(t)
UNKNOWN SYSTEM

PARAMETER Mist
W(t) ADJUSTMENT COMPUTE ake

ý(t)
MODEL
A known input, or class of input, is applied to an unknown system and to
a model that simulates the system. The difference between the two
outputs is used to fit the model, and the procedure is repeated again.
Typically the model has a fixed form, and only a finite number of
parameters need to be tuned. This procedure can be used to divide the
input signal into finite time records and perform an identification after
each of these records. If the current unknown system is too difficult for
experimentation, it should be replaced by a stored input-output log.
In applying this procedure, there are several practical problems that must
be considered. These include:
1. How to select the model structure
2. The possibilities of the time scale
3. Errors in judgment
4. Initial conditions
5. Adjustment strategy

The model selection problem is an integral part of identification methods.

The parameter tuning technique is probably the most difficult part of
the reference model method. The sources of this difficulty are:

1. Existence of multiple minimum points

2. Extreme sensitivity or insensitivity of parameters
3. Poor convergence of some models
4. Lack of orthogonality
 Adaptive Reference Model (ARM)

The adaptive control system is a system that continuously and

automatically measures the desired dynamic characteristics, and uses
the difference to vary the adjustable parameters of the system, or to
generate an actuating signal, so that optimal performance is maintained,
regardless of environmental modifications.

If a control system changes its behavior, due to changes in process

parameters, then someone must adjust it.
Adaptive Reference Model (ARM)
The general idea behind Model Reference Adaptive Control (MRAC,
also known as a MRAS or Model Reference Adaptive System) is to
create a closed loop controller with parameters that can be updated to
change the response of the system. The output of the system is
compared to a desired response from a reference model. The control
parameters are updated based on this error.

The objective is to have parameters that converge to ideal values that

cause the plant response to match the response of the reference
model.
For example
• You may be trying to control the position of a robot arm, it vibrates
naturally. You actually want the robot arm to make rapid movements
with little or no vibration. Using MRAC, a reference model can be
chosen that could respond quickly to a step input with a short
response time. You could then create a controller that would adapt
to make the robot arm move just like the model.
Stochastic Reference Model

The Stochastic reference model means being or having a random

variable. A stochastic model is a tool for estimating probability
distributions of possible outcomes by allowing for random variation in
one or more inputs over time. Random variation is generally based on
fluctuations observed in historical data for a selected period using
standard time series techniques. The distributions of potential
outcomes are derived from a large number of stochastic simulations
(projections) that reflect random variation in the input(s).
Adaptive Control by Neural Reference Model

The following example is intended to describe how the adaptive

control by reference model offered by Matlab works.
This tool uses neural networks to optimize and adapt the control
variables of our process to the variations that the system may
experience.
 Artificial Neural Networks

Artificial neural networks (ANN) are based on the functioning of the

nervous system. Neurons are grouped together to perform intelligent
tasks, organized into layers that are either supervised or
unsupervised.

Using training algorithms, the weights and thresholds are adjusted

until the error function is less than a minimum.
 Artificial Neural Networks

Why use RNA for control?

As we already know, the adaptive controller adjusts the process
parameters and/or the gains that vary due to disturbances to the
process.
By using artificial neural networks, it is ensured that stability will be
achieved in the system.

To ensure the efficiency of ANNs, the learning stage is carried out

from examples, allowing generalization and fault tolerance.
Knowledge is distributed throughout the set of neurons and their
interconnections.
The Matlab Simulink tool and neural network libraries oriented to system
control will be used, using the “Model Reference Controller” controller.

Let the equation be:

, 867200
 GS “ +10175 + 141300

We create the function in Simulink:

We create our reference model, which we assume for some reason has
a negative profit.
Now with those two diagrams we create our main diagram:
We have included the Matlab neural reference model controller, this
system being a closed loop. Then double click on this driver and we will
see the following panel:
Then clicking on the “Plant Identification” button we arrive at
this panel:

• From here we select the maximum and

minimum inputs, system interval time and
the number of hidden layers and search
for our plant model.

• To obtain a good amount of data we

choose 6000 points.

• We click on "Generate Training Data".

A graph like this will be with the configured parameters
generated above for the plant.
We select “Accept Data” and we will return to the previous window, we
set the number of epochs to 300 and click on “Train Network” this should
try to minimize the error in 300 attempts.
It will open a graph of the resulting effect, and will also show us the
parameters that we modified. We select ok, we return to the previous
panel.
Neural Network Training (nntraintool)
51 Testing data for NN Model Reference Control - | Validation data for NN Model Reference Contr...
Neural Network File Edit Yiew Insert Tools Desktop Window Help
Ie Edit Yievr Insert Iools Desktop Window Help

Layer Layer
Input Plant Output
nput Output Input Plant Output

Algorithms Da i Ih Pili minian nfLHhIm □ Training data for NN Model

Reference Control
Training: Levenberg - Marquardt (train m) Mean
Performance: Ric Edit Yiew Insert Tools Desktop Window Hep
Squared Error (mse)
Data Division:
Specified (ci eind
5 10 Input Plant Output
Progress
Mistake
Epoch: 3DD iteration 5 NN Output
1
Time: 0:0026
5.796-03 0.00834 0.00
Performance 05
Gradient: 1.00 1.00e 1C

Mu: 0.00100 -.03: । 1.00+10

0
CO
Validation Checks:
-0.5
Plots 10 2D

-1
Performance (plotperform) Err NN Output
5 10
1 5 10
Training State (plottrainstate) time (st

Regression (plotregression) 05
time [s]
Plot Interval: The
Epochs 0

-0.5
10 ch reached.

Stop reining O Cancel -1

10 20 0 10 20

time (s) time (s)

Back to the Reference Model options, we need to load the time for our
desired control in system mode. Will generate
a graph with adjusted time.

We click on “Accept Data” and returning to the Reference Model options,

100 epochs are chosen and finally we click on training controller.
After the system has been trained we obtain the error reduction graph
and another graph of inputs and outputs of the controller.
The system is now trained, as the desired system and the Neural
Controller are both linked.

The blue line (controller N) follows the desired controller (yellow line)
that represents a good result.
It should be noted that the entrance is purple. It is observed that the
system follows a ramp type input as can be noted.

As can be seen, this controller is very optimal for reducing the error in
our functions, however the training and its processing in Matlab
sometimes takes an excessive amount of time, depending on the
computing equipment and the number of epochs assigned to the
controller.