Controller for Mechatronic

Systems – An Introduction

Mr.A.Balasubramanian,
Assistant Professor / EEE,
SSN College of Engineering.
Email : [email protected]
Mr.A.Balasubramanian, AP / EEE / SSNCE 1
 ORGANIZATION
• Systems Analysis - Introduction
• Need for Controllers - Servo and regulatory mode
• Types of Controllers - Discontinuous Vs Continuous
• P, I, D - Individual Operation and Analysis
• PI, PD, PID, IP, IPD
• IO Vs FO Controllers
• Adaptive Control methods
• Application of ANN as Compensators
• Application of MLP as Controller

Mr.A.Balasubramanian, AP / EEE / SSNCE 2

 Systems Analysis - Introduction
• Watt‟s Fly-Ball Speed Controller
• Inverted Pendulum Process
• Satellite Attitude control – LAM and AOCS
• DC and AC Servomotors for Speed and
Position control
• Classical SISO and state variable MIMO
• Interaction between the subsystems

Mr.A.Balasubramanian, AP / EEE / SSNCE 3

 Mathematical Modeling
• Inverted Pendulum Equation

• Two Tank Interacting System

• Thermal system CSTR

• “RESPECT THE UNSTABLE”

• Unstable systems give deeper view to understand the importance of
control engineering, in fact, gives much opportunity to “CONTROL”.

Mr.A.Balasubramanian, AP / EEE / SSNCE 4

 State equations DC machine (sep.Exc.)
dia ra Kv 1
 ia  r  Va
dt Laa Laa Laa
dr Bm Kv 1
 r  ia  TL
dt J J J

X  AX  BU
 ra Kv   1 
   0 
 a
i
P     aa L Laa   ia    Laa Va 
    
 r   K v 
Bm   r  
0 
1  TL 
 J J   J 
Mr.A.Balasubramanian, AP / EEE / SSNCE 5
 Need for Controllers D(s)

• Came into use in 1930‟s with the

Gd(s)
introduction of pneumatic controllers.
• Extremely flexible and powerful control
Ysp(s) E(s) C(s) U(s) Y(s)
algorithm when applied properly. +- Gc(s) Ga(s) Gp(s) ++
• However, it was probably not until
Minorsky‟s work on ship steering*
Ys(s)
published in 1922, that rigorous Gs(s)
theoretical consideration was given to
PID control, the first mathematical
treatment of the type of controller.

Mr.A.Balasubramanian, AP / EEE / SSNCE 6

 P,PI control- Introduction
• P Control
• In proportional control there is a continuous linear relation between the value of the
controller output and the error. This relation may be given as
•
• u(t) = Kp e(t) + uo

• where u(t) is the controller output at time t, e(t) is the error at time t, Kp is the
proportional gain and uo is the controller output for zero error.
• PI Control
• In PI Control, the controller output is proportional to the error and integral of error.
This relation is given by
•
• u(t) = Kp { e(t) + KI e(t) dt } + u(0)

• where KI is the integral gain and u(0) is the controller output for t=0.

Mr.A.Balasubramanian, AP / EEE / SSNCE 7

 PID control
• PID Control
• In PID Control, the controller output is proportional to the error, integral of
error and derivative of error. This relation is given by
• e(t)- the error from set point [e(t)=ysp-ys].
• The controller gain is a tuning parameter and largely determines the
controller aggressiveness; The reset time is a tuning parameter and
determines the amount of integral action; The derivative time is a tuning
parameter and determines the amount of derivative action.

 1
t
de(t ) 
c(t )  c0  K c e(t )   e(t )dt   D 
  I 0 dt 
where e(t )  ysp  ys (t )

• Rate limiter are used to avoid

derivative kick in the output.

Mr.A.Balasubramanian, AP / EEE / SSNCE 8

 Example for a First Order Process with
a PI Controller
• Overdamped | Underdamped
K c  1;  I  1; K p  1;  p  5;   2
Kc  2  I  10 Kp 1 p 5
Characteri stic Equation :
Characteri stic Equation :
 1  1
 1  2   25s 2  20s  1 1    1  0
  2    1 0   s
 5s  1   10 s 
Rearrangin g
Rearrangin g
25s 3  20s 2  2 s  1  0
25s 2  15s  1  0
p1   0.764 and a second order
 p  5   1.5 response with  p  4.37 and   0.08

Mr.A.Balasubramanian, AP / EEE / SSNCE 9

 ZN and CC method of tuning of PID controller

Continuous oscillation method Reaction Curve method

Cohen and Coon (CC)carried out further studies to find controller settings
which, based on the same model, lead to a weaker dependence on the
ratio of delay to time constant(better than over Ziegler-Nichol‟s method).

Mr.A.Balasubramanian, AP / EEE / SSNCE 10

 Digital Equivalent of PID Controller
• The trapezoidal approximation of the integral.
• Backward difference approximation of the first derivative
 n d e(t ) e(t )  e(t  t )
0
e(t ) dt   e(i t ) t
i 1
dt

t
 t n e(t )  e(t  t ) 
c(t )  c0  K c e(t ) 

 e(i t )   D


 I i 1 t 
 t n 1 e(t  t )  e(t  2t ) 
c(t  t )  c0  K c e(t  t )   e(i t )   D 
  I i 1  t 
________________________________________________________
 t e(t )  e(t )  2e(t  t )  e(t  2t ) 
c(t )  K c e(t )  e(t  t )   D  
 I  t 

Mr.A.Balasubramanian, AP / EEE / SSNCE 11

Three versions of PID algorithm
• The original form in which the proportional, integral, and derivative terms
are based on the error from set point
• The form in which the proportional and integral terms are based on the
error from set point while the derivative-on-measurement is used for the
derivative term.
• The form in which the proportional and derivative terms are based on the
process measurement and the integral is based on the error from set
point.( All these three forms are used in DCS‟s.)
 t e(t )  e(t )  2e(t  t )  e(t  2t ) 
c(t )  K c e(t )  e(t  t )   D  
  I  t 

 t e(t )  y (t )  2 ys (t  t )  ys (t  2t ) 
c(t )  K c e(t )  e(t  t )   D  s 
 I  t 

 t e(t )  y (t )  2 ys (t  t )  ys (t  2t ) 
c(t )  K c  ys (t  t )  ys (t )   D  s 
  I   t 

Mr.A.Balasubramanian, AP / EEE / SSNCE 12

 Adaptive control Techniques
• GS,MRAC,STR
• MIT rule

Mr.A.Balasubramanian, AP / EEE / SSNCE 13

 „IO‟ vs „FO‟ Controllers
• Response of different controller configurations
where curves 1, 2, 3 and 4 correspond to FO-PID,
FO-PI, PI and PID respectively. It is understood
from the closed loop response that the FO
controllers are more advantageous in their
performance than the integer order controllers in the
electro-dynamic levitation applications.

Mr.A.Balasubramanian, AP / EEE / SSNCE 14

 Supervised control
• A neural network learns
the mapping from sensor
inputs to desired actions
by adapting to a training
of examples of what it
should do.
• An automatic controller
mimics the action of the
human - hence this
technique is called
supervised control.

Mr.A.Balasubramanian, AP / EEE / SSNCE 15

 Direct inverse control
• Neurocontroller B is an exact copy of neural network A,
which undergoes training. Neural Network A is
connected in such a way that it gradually learns to
perform as the unknown plant inverse.
• The input of the controller is d which is the desired
response of the plant.
• The plant‟s actual response is y due to its input x
produced by the neurocontroller B. The error used for
training the neural network is the difference between the
output signals of networks A and B.
• Although the network B tracks A after each training step,
Y, must exactly match d for this error to reduce to zero.
• This control architecture is also called an indirect
learning architecture.
Mr.A.Balasubramanian, AP / EEE / SSNCE 16
 Feed forward Control with plant inverse
learning
• This
architecture
can be used
for IMC for
systems with
non-minimal
zeroes also

Mr.A.Balasubramanian, AP / EEE / SSNCE 17

 Direct Adaptive control
• The control objective of specialized learning control
architecture is to obtain the desired output d of the plant
under condition that the plant input causing this output is
unknown.
• For a dynamical plant, the trained neural network needs
to be provided the sequence of recent output plant
vectors in addition to the present plant output. The
control objective of specialized learning control
architecture is to obtain the desired output d of the plant
under condition that the plant input causing this output is
unknown.
• For a dynamical plant, the trained neural network needs
to be provided the sequence of recent output plant
vectors in addition to the present plant output.
Mr.A.Balasubramanian, AP / EEE / SSNCE 18
 Online learning Architectures

• Static and Dynamic

Processes.

Mr.A.Balasubramanian, AP / EEE / SSNCE 19

Comparision of Computing Approaches

• Qualitative analysis

Mr.A.Balasubramanian, AP / EEE / SSNCE 20

Expert systems vs ANN

Mr.A.Balasubramanian, AP / EEE / SSNCE 21

 ANN based DZ Pre-compensators
• Physical Nonlinearity
Compensators.
(Hammerstein‟s NL)
• Dead Zone

Mr.A.Balasubramanian, AP / EEE / SSNCE 22

 Controller with ANN Pre-compensator

• Robotic Arm
position control
system.
• Tracking
controller tuned
with PD
controller and
ANN (M-MLP)
precompensator

Mr.A.Balasubramanian, AP / EEE / SSNCE 23

 Performance Analysis
• Normal PD
control gives
satisfactory
result without
the
Nonlinearity
• With Nonlinear
element, the
tracking
property is
poor.
• With ANN Pre-
Compensator,
tracking is
better and
relative
stability is
improved.

Mr.A.Balasubramanian, AP / EEE / SSNCE 24

 ANN based Backclash Pre-compensators

• One of the
nonlinearities in
mechanical system
• Estimation of inverse
function is a tedious
one.

Mr.A.Balasubramanian, AP / EEE / SSNCE 25

 Backclash Compensator performance
• Mechanical system tuned with
PD controller and with ANN
(M-MLP) pre-compensator

Mr.A.Balasubramanian, AP / EEE / SSNCE 26

 Response equation for speed control
 1  1  1
  a m Va    P  TL
 Kv   J  a 
r 
 1 Bm   1  1 Bm 

P  
2

 P     
 a J    a   m J 
 
2  n2

n P 2  2P   n2  0
G1  with TL  0 ( No - Load Condition)
Va   exponential decaying factor
n  n  undamped natural frequency
G2  with Va  0
TL oscillatio n of the sytem

  damping factor
n
b1,b2   n   n  2  1

Mr.A.Balasubramanian, AP / EEE / SSNCE 27

 Equations for speed control
• Response equation

• Discretised for a
sampling period T.

• K1 to K6 are the
constants to be
evaluated from the
machine and load
parameters.

Mr.A.Balasubramanian, AP / EEE / SSNCE 28

 ANN Controller-Points to remember
• Data collection for training and Testing (PRBS and MLS
more suited than the conventional standard test signals)
• ANN Architecture( selection of hidden layers and
neurons with or without threshold-data length)
• Checking for the excessive filtering ability.
• ANN parameters and training algorithm.
• Off line testing and Online implementation.
• Computational delay and checking the degeneracy
condition.
• Satisfying the performance index ISE,ITAE etc..

Mr.A.Balasubramanian, AP / EEE / SSNCE 29

 Simulation and analysis
net.trainparam.epochs=10; net.trainparam.epochs=50;
net.trainparam.goal=1e-4; net.trainparam.goal=1e-6;
30 30

20 20

10 10

0 0

-10 -10

-20 -20

-30 -30
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500

plot(1:kmax,wr,'b',1:kmax,w_sim,'r')

Mr.A.Balasubramanian, AP / EEE / SSNCE 30

 Queries
• Any questions?

.. Dedicated to the memories of my

parents and teachers ( Mathru devo Bhavaa;
Pithru devo Bhavaa; Aacharya devo Bhavaa.)
who gave me a lot and asked nothing…
… but my growth only.

Yours Sincerely,
ABS / AP / EEE / SSNCE

Mr.A.Balasubramanian, AP / EEE / SSNCE 31