500 likes | 952 Views
NEURO - FUZZY CONTROL A CASE STUDY. DR. T. THYAGARAJAN PROFESSOR & HEAD DEPT. OF INST. ENGG. ANNA UNIVERSITY, MIT CAMPUS thyagu_vel@yahoo.co.in. CONTENTS. What is FLC? Where FLC? Components of FLC Applications Advantages Case study Disadvantages NFC design
E N D
NEURO - FUZZY CONTROL A CASE STUDY DR. T. THYAGARAJAN PROFESSOR & HEAD DEPT. OF INST. ENGG. ANNA UNIVERSITY, MIT CAMPUS thyagu_vel@yahoo.co.in
CONTENTS • What is FLC? • Where FLC? • Components of FLC • Applications • Advantages • Case study • Disadvantages • NFC design • Closed loop studies • Comparison of performance indices • Conclusion • Future Scope
WHAT IS FLC? • FLC emulates the human mind for monitoring the process parameters and takes decisions regarding the control action • FLC converts a linguistic control scheme using expert knowledge base into an automatic control stratergy
WHERE FLC? • Where one or more variables are continuous • Where mathematical model of the process does not exist (or) too complex to evaluate the model • Where high ambient noise level has to be dealt with • Where inexpensive sensor / low precision microcontroller are to be used • Where the expert knowledge about the system behaviour is available
ADVANTAGES OF FLC • Detailed mathematical model is not necessary • Ideal for complex/nonlinear systems • Compatible with existing control system • Provides robust control • Hardware implementation is possible • Demonstrates smooth control action even with small number of rules
APPLICATIONS • REFRIGERATOR • AIR CONDITIONER • WASHING MACHINE • VIDEO CAMERA • HOT WATER HEATER • LIFTS/ELEVATOR • ELECTRIC TRAIN • PROCESS/SYSTEM CONTROL
COMPONENTS OF FLC • Fuzzification • Knowledge base • Decision making logic (or) inference engine • Defuzzification
Fuzzification • Measure the input variables (error, change in error/integral error) • Convert the input variables into suitable linguistics values (VS = Very Small, S = Small, M = Medium, L =Large, VL = Very Large etc) Convert the input variables into corresponding universe of discourse using membership function)
KNOWLEDGE BASE • (a) Data Base • (b) Rule Base • Data base is used to define linguistic control variables • IF <fuzzy proportion > THEN <fuzzy proportion > • ‘IF’ part is called ‘antecedent ‘(e,ce,ie) • ‘THEN’ part is called ‘consequent’ (mv) • The combination is called ‘premise’
DECISION MAKING LOGIC (OR)INFERENCE ENGINE • Capability of simulating human decision making process • Infers a system of rules through fuzzy operators namely ‘AND’ and ‘OR’ • Generates a single truth value using Max-Min criteria
DEFUZZIFICATION • Yields a crisp, non-fuzzy control action. • (i) Max-Criteria • (ii) Mean of the maximum • (iii) Centre of area method Z0 = j . X j • -------- i N is the number of quantization levels j is the max. value of membership corresponds to ith quantization level X j is the support value at which membership function reaches maximum value
DESIGN OF OPTIMAL PID CONTROLLER • MODEL OF THE AHS AS FOPDT • USE Z-N TUNING RULE TO FIND THE INITIAL PID CONTROLLER SETTINGS • BY TRIAL-AND-ERROR TUNE THE PID CONTROLLER FOR OPTIMAL SETTINGS • FIND e(t), ie(t) or ce(t) and m(t) and use them as knowledge base
Design details of PID • AHS FOPDT MODEL = (0.2 X e -16s)/(1 +220 s) Optimal PID controller settings: Kc =67.56 Ti =31.4 Td =7.85
DESIGN OF TRADITIONAL FLC • Input variables = e(t) and ie(t) • Quantization levels = 5 • For e(t): MN,N,Z,P and MP • For ie(t): VS,SM,L and VL • For u(t): VS,SM,L and VL • Membership function: Triangular • Truth value generation: Max-min Criteria • Defuzzification: Centre of area method
FUTURE SCOPE • Hybrid control strategies can be designed using FLC, ANN and GA • ANN can be used to generate the membership values • GA can be used to tune the FLC • Adaptive FLC • Neuro fuzzy control
Neuro-fuzzy control- case study • In the conventional fuzzification, finding the corresponding universe of discourse value for every quantization level needs repeated computation. • In the case of NFC, the conventional fuzzification is replaced by ANN technique • Two ANN models, one for e(t) and other for ie(t) are formulated • These ANN models are used to carryout fuzzification
ANN parameters for e(t) • Input neurons: 2 • Hidden neurons: 1 5 • Output neurons: 5 • Bias: 1 • Learning rate: 0.7 • Momentum factor: 0.3 • Iterations: 31,600
ANN parameters for ie(t) • Input neurons: 2 • Hidden neurons: 1 2 • Output neurons: 5 • Bias: 1 • Learning rate: 0.7 • Momentum factor: 0.3 • Iterations: 25,000
Conclusion • ANN based fuzzification avoids the repeated computations carried out in the conventional fuzzification. • Robust ANN models for e(t) and ie(t) can be formulated with minimum number of input-output data pair • The iterations required for convergence are also less