Neural Network Application in

Intelligent control
February 09/2021
 Application of NN
• Neural network for modeling and estimation
– Plant model /state observer
– Inverse model
• Neural network for control
– Internal model control
– Model predictive control
– Smith predictor control
– Model reference control
• Implementations
– General purpose hardware
– Special purpose hardware
AAIT - SECE- Industrial Control Stream
04/11/2021 2
Dr. Dereje Shiferaw
 1. Neural network for modeling and
estimation
• ANN can model any system
– Linear
– Nonlinear
• Limitation
– Availability of input-output data

AAIT - SECE- Industrial Control Stream

04/11/2021 3
Dr. Dereje Shiferaw
 What is a Model?
• Model
– Meaning
• representation of construction and working of system,
it is similar to but simpler than the system it represents
– Purpose
• to enable the analyst to predict the effect of changes to
the system
– Good model
• close approximation to the real system
• incorporate most of its salient features
• Not so complex
• judicious tradeoff between realism and simplicity

AAIT - SECE- Industrial Control Stream

04/11/2021 4
Dr. Dereje Shiferaw
 1.1 Plant model
• NN is used to model input-output relations of
a dynamic system(plant)
• Steps in plant modeling
– Collect Plant input-output data
– Pre-processing
– Determine NN architecture
– Train network
– Test the network

AAIT - SECE- Industrial Control Stream

04/11/2021 5
Dr. Dereje Shiferaw
 NN used for plant model
• Block diagram representation

Plant

NN

AAIT - SECE- Industrial Control Stream

04/11/2021 6
Dr. Dereje Shiferaw
Example- model of a heat exchanger system

• Objective – is maintain liquid temperature by

controlling steam flow rate

AAIT - SECE- Industrial Control Stream

04/11/2021 7
Dr. Dereje Shiferaw
 Step 1- Measured input output data
Time 0 20 30 40 50 60 70 80 100
Tank T 0.0113 0.196 0.559 0.705 0.774 0.922 0.921 0.985 1.05

Step 2 preprocessing of data

total sample data = 9
training data =7
test data =1, [40 0.705]
validation data=1 , [80 0.985]

AAIT - SECE- Industrial Control Stream

04/11/2021 8
Dr. Dereje Shiferaw
Step 3- determine NN architecture
• Let us assume that FF with 1 hidden layer is to
be used
• 1-5-1 structure, sigmoid activation for hidden
layer and linear activation for output layer

AAIT - SECE- Industrial Control Stream

04/11/2021 9
Dr. Dereje Shiferaw
 Step 4- create network and train it
• net=newff(minmax(datat),[5 1],{‘tansig’,
’purelin’});
• net=train(net,datat,ouput);

AAIT - SECE- Industrial Control Stream

04/11/2021 10
Dr. Dereje Shiferaw
 AAIT - SECE- Industrial Control Stream
04/11/2021 11
Dr. Dereje Shiferaw
 Step 5-test and validation
• Sim(net,40)=0.7116
• Eroor=0.705-0.7116=-0.0066
• Sim(net,80)=0.9373
• Error=0.985-0.9373=0.05

AAIT - SECE- Industrial Control Stream

04/11/2021 12
Dr. Dereje Shiferaw
Example 2- dc motor dynamic modeling

• Plant – linear plant with fourth order TF

• Input output data is collected for step input of
0.1 to 1 in step of 0.1
• 200 sample data values are used
• NN used is FF with two layers

AAIT - SECE- Industrial Control Stream

04/11/2021 13
Dr. Dereje Shiferaw
 1.2 NN used as state observer
• State observer
– Estimates the state of the system
– Example – in DC motor state may be either
[velocity and position] or [current and velocity]
– Robotics – estimates inverse or forward
kinematics

AAIT - SECE- Industrial Control Stream

04/11/2021 14
Dr. Dereje Shiferaw
 NN state observer
AAIT - SECE- Industrial Control Stream
04/11/2021 15
Dr. Dereje Shiferaw
 1.3 Inverse model
• A plant
– Control signal/manipulated variable as input
– Controlled variable as output
• NN in direct modeling or observer
– Input- control signal
– Output is estimated state/output
• NN used in Inverse model
– Input is output of plant
– Output is estimated control signal
AAIT - SECE- Industrial Control Stream
04/11/2021 16
Dr. Dereje Shiferaw
 Training NN for inverse model

AAIT - SECE- Industrial Control Stream

04/11/2021 17
Dr. Dereje Shiferaw
 2. NN used as controller
• 2.1 replacing existing controller
• 2.2 Internal model control
• 2.3 smith predictor
• 2.4 model predictive control
• 2.5 model reference control

AAIT - SECE- Industrial Control Stream

04/11/2021 18
Dr. Dereje Shiferaw
 2.1 Replacing existing controller
• Assume there is an existing controller
• Error input and control signal output of
existing controller is collected
• NN is trained off line
• Used to replace controller

AAIT - SECE- Industrial Control Stream

04/11/2021 19
Dr. Dereje Shiferaw
• Replacing existing controller

AAIT - SECE- Industrial Control Stream

04/11/2021 20
Dr. Dereje Shiferaw
 2.2 Internal model control(IMC)
• Consider the following block diagram

• G(s) is plant, Gm(s) is nominal model, R(s) is reference , U(s) is

control signal, D(s) and N(s) are disturbance and noise signals
respectively
AAIT - SECE- Industrial Control Stream
04/11/2021 21
Dr. Dereje Shiferaw
 Analysis of IMC
• From the block diagram

• Sensitivity function is given by

AAIT - SECE- Industrial Control Stream

04/11/2021 22
Dr. Dereje Shiferaw
 IMC analysis contd…
• Complementary sensitivity

• Perfect tracking
Gm(s)=G(s) and C(s)=1/G(s)
• Perfect disturbance rejection
Gm(s)=G(s) and C(s)=1/Gm(s)
AAIT - SECE- Industrial Control Stream
04/11/2021 23
Dr. Dereje Shiferaw
 NN application in IMC
• Conditions for perfect set point tracking and
disturbance rejection are
– Gm(s)=G(s) and C(s)=1/G(s)
• NN model is used for Gm(s)

AAIT - SECE- Industrial Control Stream

04/11/2021 24
Dr. Dereje Shiferaw
 2.3 Smith predictor
• Smith predictor is similar to IMC
• Used for systems with time delay
• Time delay limits performance of conventional feedback
– Adds phase lag and affects stability
– Gain must be reduced
– Response becomes sluggish
• Sources of time delays
– Distance velocity lags
– Recycle loops
– Time for analysis of compositions

AAIT - SECE- Industrial Control Stream

04/11/2021 25
Dr. Dereje Shiferaw
 Time delay compensation
• Time delay limits performance of conventional
feedback
– Adds phase lag and affects stability
– Gain must be reduced
– Response becomes sluggish
• Sources of time delays
– Distance velocity lags
– Recycle loops
– Time for analysis of compositions
AAIT - SECE- Industrial Control Stream
04/11/2021 26
Dr. Dereje Shiferaw
 Smith predictor
• The plant G is modeled by a system G*(s) and
delay

~ is used as feedback
• The predicted output Y
AAIT - SECE- Industrial Control Stream
04/11/2021 27
Dr. Dereje Shiferaw
 Analysis of Smith predictor
• From the model

• For perfect modeling

G*(s)=G(s) and time delay should be
correctly predicted , then

• NN is used as G*s(s)
AAIT - SECE- Industrial Control Stream
04/11/2021 28
Dr. Dereje Shiferaw
 2.4 Model predictive control
• Objective
– Control MIMO system having in equality
constraints
• Problem formulation
– Given an MIMO system and its model, predict
future outputs using model and present inputs
and compute the necessary changes in input

AAIT - SECE- Industrial Control Stream

04/11/2021 29
Dr. Dereje Shiferaw
 MPC contd…
• Advantages
– Model captures static and dynamic interaction
between input/output and disturbance
– Constraints are considered
– Control calculations are coordinated with
optimum set points
– Accurate prediction warns potential damage

AAIT - SECE- Industrial Control Stream

04/11/2021 30
Dr. Dereje Shiferaw
 MPC block diagram

AAIT - SECE- Industrial Control Stream

04/11/2021 31
Dr. Dereje Shiferaw
 MPC contd…
• Objectives of MPC
– Prevent violations of input output constraints
– Drive some output variables to their optimal set
points while maintaining other output with in
specific range
– Prevent excessive movements of input variables
– Control as many process variables

AAIT - SECE- Industrial Control Stream

04/11/2021 32
Dr. Dereje Shiferaw
 Model reference control
• The Model-Reference Adaptive system
(MRAS) was originally proposed to solve
a problem in which the performance
specifications are given in terms of a
reference model.
• This model tells how the process output
ideally should respond to the command
signal.
• The Adaptive Controller has two loops.
The inner loop consists of the process
and an ordinary feedback controller.
AAIT - SECE- Industrial Control Stream
04/11/2021 33
Dr. Dereje Shiferaw
 MRAC Introduction contd…
• The outer loop adjusts the controller parameters
in such a way that the error, which is the
difference between the process output y and
model output ym is small.
• The MRAS was originally introduced for flight
control.
• In this case, the reference model describes the
desired response of the aircraft to joystick
motions.
• Model reference adaptive systems were originally
derived for deterministic continuous systems.
AAIT - SECE- Industrial Control Stream
04/11/2021 34
Dr. Dereje Shiferaw
 AAIT - SECE- Industrial Control Stream
04/11/2021 35
Dr. Dereje Shiferaw