Direct adaptive control
Reference y
r Model M
- e
ec
u Plant +
Controller
P
yp
Indirect adaptive control
Reference ym
Model M
ei - ec
y
r Identification +
Model - ei
ec
u Plant yp +
Controller
P
The simulation and results of the given non-linear dynamical system-1 are taken.
The dynamics of the system is:
yp(k+1) =f [yp(k),yp(k-1)] +u(k)
f [yp(k),yp(k-1)] = yp(k) yp(k-1) [yp(k)+2.5]
1+ yp2(k)+ yp2(k-1)
yr(k+1) = 0.6 yr (k) + 0.2 yr (k-1) + r(k)
r(k) = sin(2k/25) or 5
u(k) = -f [yp(k),yp(k-1)] + 0.6 yp (k) + 0.2 yp (k-1) + r(k)
u(k) = -N [yp(k),yp(k-1)] + 0.6 yp (k) + 0.2 yp (k-1) + r(k)
yp(k+1) =f [yp(k),yp(k-1)] -N [yp(k),yp(k-1)] +0.6 yp (k) +0.2 yp(k-1) + r(k)
Output of the Direct Adaptive Controller using
the sinusoidal reference signal
The output of the controller (u) and the system output
Output of the Indirect Adaptive Controller using
the constant reference signal
The output of the controller (u) and the system output
Most of the reported
applications are still in
research stage
No formal proofs, but they
seem to have useful
applications that work
STRUCTURE OF A NEO FUZZY NEURON IN SYSTEM
IDENTIFICATION
REFERENCE
MODEL
YP(k)
FUZZY NEURAL
NETWORK
X(k) Y(k)
(k,i) e(k)
LEARNING
ALGORITH
M
ADVANTAGES OF NEURO FUZZY TECHNIQUES
• Non model based
• Real time implementation
• For large ranges of data ,scaling as in case of NN is not
required,due to fuzzification of inputs
• Increased accuracy in identification & control as significance of
each input is known
• Antecedents of all rules are described by complementary
membership functions ,so that defuzzification is achieved
without arithmetic division
NEURO- FUZZY MODEL
Each input is Fuzzified ,all consequents are singletons,defuzzification
is done at synapse.
Input matrix X =[x1,x2,………,xp,…………,xn]; n = total number of
inputs.
This input matrix is fuzzified using any gaussian membership function.
(k,i) is the membership function of the fuzzy set, which is input of
the NN.
n n
(k , i) ( x p ) exp([a pi {x p c pi }]lpi )
i
p
p 1 p 1
where,
api determines the width of membership function,
cpi represents the center of membership function and
lpi give the shape of membership function
If Then Rule : IF xi lies in the fuzzy segment µij, THEN the
corresponding weight wij should be increased directly proportional to
output error (e(k)), because error is caused by the weight.
wij(k+1)=wij(k)+ai.µij(xi)e(k)
The output of neural defuzzifier y(k) is given by the following equation:
r
y (k ) { w(k , i ) (k , i ) b}
i 0
where,
b is the threshold value ;
is non-linear activation function;
w are the weights of the neural network .
Identification using neuro fuzzy neuron is done;on line and off line
control is done using NN controller
LEARNING AND ADAPTIVE ALGORITHM
Mean square of error (MSE) minimization technique for learning and adaptation of
weights is used.
For this we choose a performance index (J) for the system
J 0.5e 2 where e=yp -y
So that the weight update equation is :
wnew wold w where J
w
w
J J e y
(For Plant Identification)
w e y w
J J ec y u
(For Control of plant)
w ec y u w
is step size , u is output of the neural controller.
NEURO FUZZY IDENTIFICATION
PREDICTION
DIRECT CONTROL
IN-DIRECT CONTROL
DISCUSSION OF
RESULTS
• IDENTIFICATION
Performance index minimizes at 2nd iteration
• CONTROL
Direct control:
Square of error minimizes at 10 th iteration.
Indirect Control:
Square of error reduces at 10th iteration.
COMPARITIVE STUDY FOR IDENTIFICATION
IDENTIFICA- NEURAL NEURO
TION NETWORKS FUZZY
PERFORMA-
NCE INDEX 85 6.155
AT 35TH
ITERATION
COMPARITIVE STUDY FOR CONTROL
SQUARE OF
ERROR AT NEURAL NEURO
TENTH NETWORKS FUZZY
ITERATION
DIRECT 60 0.506
CONTROL
INDIRECT
CONTROL 40 0.50
CONCLUSION
• Neuro-fuzzy shows better performance over neural and fuzzy
techniques in system identification.
• In-direct or on-line control gives better results than direct or
off line control.
DISCUSSION OF RESULTS
1. IDENTIFICATION
From the simulation result for the performance index it is
observed that there is a local minima after five iterations; which
increases sharply which decays at hundredth iteration.
Minimum value of performance index is thirty six at one hundred
and sixth iteration; thereafter remaining almost constant.
2. CONTROL
Direct Control:
Square of error remains almost constant for all values of
iterations except at one hundredth iteration where it rises sharply.
Minimum value of square of error is fifty at tenth iteration.
Indirect Control:
After a sharp peak square of error lowers down at fifteenth
iteration and minimizes to ten after twentieth iteration.
A Neo -Fuzzy neuron and a series of delay
elements used in identification
CONTROL
Direct adaptive control
Indirect adaptive control
Identification as well as control using Neural
Network
Implementation of Neural Network Controller on
Real Time System
Comparison Between Neural Network and PID
Controllers
• Modeling And Identification
of Plant
• Design And Simulation of
NN Controller
• Implementation of Designed
Controller On Real Time
System
• Comparison of Designed
Controller With PID
Controller As Reference
Tank1- Infinite Source of Compressed air Tank2- Reservoir
CV1- Pneumatic Control Valve V2- Disturbance
E/P Converter Pressure Transmitter
Basis of Modeling
Priori Knowledge- Not known in our case
Input Output data- Acquired using interface through MATLAB and
Data Acquisition Cards
Input- Signal Voltage (0-5V)
Output- Pressure in Tank 2
Basis of Identification
Learning Rules
Prediction
IDENTIFICATION
TARGET
INPUTS
NEURAL NETWORK
COMPARE
OUTPUT
ADJUST
WEIGHTS
NN- Feed Forward with
single hidden layer
Learning Rule-
Gradient Descent
Algorithm
Training Rule-Back
Propagation
The Equations for the output of the NN are:
yi ( P * w1 b1 ) y ( yi * w2 b2 )
For fine tuning of parameters, an objective function is
chosen
1 M 2 where, M is the total number of
J e ( j)
2M j 1 samples
Parameter Update Formulae
wnew = wold + w
J J e y yi
w1 2 2 2 e(1) w2 P
w1 e y yi w1 yi w1
J J e y
w2 2 2 2 e(1) yi
w2 e y w2 w2
ADAM – 4014D-A/D Converter
ADAM – 4520- converts RS-485 link to RS-
232 link
and vice-
versa
PCL – 726- sends the output voltage from the
controller to the final control element (control
valve).
Construct the serial port object
Connect serial object to the serial port
Receive data from ADAM – 4014D
Disconnect the serial object from the serial
port
0-5 V
Pictorial Representation:
Pressure PCL-726 NN Controller
Transmitter ADAM 4014D
RS-485 4-20 mA
ADAM 4520
E/P Converter
RS-232
Receiving The PC COM1 Sending the
Value of Output 3-15 Psi
output value of
Pressure through the NN controller
Serial to the
Communication control valve
NN Controller Control Valve
Performance Parameter- Square of Error
Controlled Variable-Output Pressure of the
Plant
Manipulated Variable- Input Voltage to the
Control Valve
NN used- Feed Forward
Learning Algorithm- Gradient Descent
Offline control
Online control
Control with Disturbance and Perturbation
from 200 to 210 samples
Simulation of PID based Controller
Simulation of PID based Controller with disturbance
and parameter perturbation from 200-210 samples
Simulation of NN based Controller
Simulation of NN based Controller with disturbance
and parameter perturbation from 200-210 samples
Real time implementation of PID controller
Real time implementation of PID controller
with disturbance and parameter perturbation
NN
Real time implementation of NN controller
NN
Real time implementation of NN controller
with disturbance and parameter perturbation
Controllers Square of Error Square of Error
with perturbation
and disturbance
PID 3.65 x 10-3 3.81 x 10-3
(Simulation)
PID 3.72 x 10-3 3.91 x 10-3
(Real Time)
NN 1.42 x 10-3 1.81 x 10-3
(Simulation)
NN 1.48 x 10-3 1.89 x 10-3
(Real Time)
Real time systems can generate outputs according
to preset value using the NN controllers.
Neural Network control adapts itself to the
nonlinearities in the system unlike PID.
Neural Network controller strategy is robust
enough with parameter perturbation and disturbed
input signal.
