Priya R et al.

/ Indian Journal of Computer Science and Engineering (IJCSE)

DESIGN OF AN ADAPTIVE
CONSTRAINED BASED NEURO-FUZZY
CONTROLLER FOR FAULT
DETECTION OF A POWER PLANT
SYSTEM
Priya R
Research Scholar, University of Kerala,
Trivandrum, Kerala 695581, India
[email protected]

Elizabeth Sherly
Professor, Indian Institute of Information Technology and Management-Kerala,
Trivandrum, Kerala 695581, India
[email protected]
Abstract: This paper proposes an adaptive constraint based framework for fault detection of a complex
thermal power plant system. In many complex systems, representation of precise and crisp constraints
uses formal specification languages such as Object Constraint Language (OCL). Here, a constraint based
neuro-fuzzy controller to tackle imprecise constrained objects is proposed. The proposed inference
system is used to identify the intensity of faults by mapping instance values against the constraints. The
imprecise constraints defined as fuzzy constraints are prioritized using fuzzy weights assigned to rules.
Back-propagation algorithm is used to train and calibrate the controller to capture the dynamic behavior
of the system. The system is adopted into a coal-fired Thermal power plant system, which is controlled by
different parameters in a constrained environment. The output of the neuro-fuzzy inference system is
compared against actual plant site alarm that functions based on rule based system and has been found
that the intensity of faults can be accurately determined.
Keywords: Fuzzy inference systems; fuzzy control; neuro-fuzzy systems; process control; fault detection and
diagnosis.
I. Introduction
Unified Modeling Language (UML) is widely used for structural modeling of any complex object oriented
system, with the help of Object Constraint Language( OCL). OCL is a powerful formal specification language
that represents constraints like invariants on classes and types, preconditions and postconditions on operations.
However, representing all the constraints of a complex system in OCL requires a rule base system that is tedious
and exhaustive. Moreover, the constraints need to be very precise to be represented using OCL. For
incorporating imprecise constraints in an intelligent manner, fuzzy control technique can be used. If the
constraints of the complex system are of different priority levels, they are mapped to weighted fuzzy rules of the
fuzzy control system.
Recently, the integration between fuzzy logic and neural network known as neuro-fuzzy or the fuzzy neural
network (FNN) has drawn attention of many researchers in the field of control engineering [1]. Combining the
learning ability of neural network and human-like reasoning of fuzzy logic makes FNN a very flexible
intelligent control technique for many control applications. In this paper, a model is proposed for detection of
faults of a constrained complex system, whose constraints are represented as fuzzy rules. The scopes for
adopting artificial intelligent based modeling for fault detection in control systems are to increased accuracy and
reliability. The proposed neuro-fuzzy model is adapted to the Boiler subsystem of a Coal-fired Thermal Power
Plant that helps to achieve generating the rule parameters automatically using neural network. Hence, constraint
representation of complex systems using a neuro-fuzzy controller doubles the advantages of the cognitive
thinking capability of fuzzy logic and learning capability of neural networks. The rest of the paper is organized
as follows. Section II gives a brief summary on the various works carried out using neuro-fuzzy soft computing
technique. Section III explains the methodology of the proposed work. The methodology is tested on a complex
system of Thermal power plant and it is discussed in Section IV. The performance of the model is evaluated and
the work is concluded with scope for future improvements in Section V.

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 208

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

II. Literature Survey

Gurpreet S. Sandhu and Kuldip S. Rattan explained how to design and train a proportional neuro-fuzzy
controller. The architecture and the learning algorithm of a proportional fuzzy controller (PFLC) is presented. A
step by step algorithm for the off-line training of a PFLC is demonstrated [2]. A survey on Neuro-fuzzy systems
was done by José Vieira, Fernando Morgado Dias and Alexandre Mota. The article summarizes the architecture
of popular neuro-fuzzy systems.[3]. Disha Sharma has explained the fuzzy modeling problem and the steps
involved in it. The paper provides an introduction to the use of fuzzy sets and fuzzy logic for the approximation
of functions and modeling of static and dynamic systems [4]. A neuro-fuzzy designer control algorithm using
neuro-fuzzy identifier is proposed by Minho Lee et al. which could be used for modelling fuzzy system whose
rules are uncertain.[5]. The authors in [6] develop a fuzzy control algorithm for a warm water plant. There are
many other applications[7] of conventional fuzzy control, including robot [8], steel furnace [9], waste water
treatment [10] and so on.[11],[12],[13],[14],[20]. In addition, fuzzy control has been widely used in various
consumer electronic devices such as video cameras, washing machines, TV, and sound systems in the late
1980sand early 1990s [15], [16]. The authors in [17] have proved that input parameter selection is very critical
in ANFIS learning.
The above literature enforces the fact that neuro-fuzzy control systems have been widely used in a variety of
control applications for improved performance. In this work, fault is detected from the model of a constrained
complex system with imprecise constraints, by modeling constraints as fuzzy rules. The model is trained for
accuracy by incorporating the learning capability of neural network.
III. An adaptive Fuzzy logic controller for fault detection of a constrained complex system
Fuzzy modeling of complex control systems without prior knowledge of fuzzy rules often requires fine tuning of
fuzzy controller. The learning capability of neural network facilitates fine tuning of design parameter of fuzzy
logic. Weights are assigned to the fuzzy rules based on the priority of constraints of the system. The proposed
fuzzy model helps to detect fault intensity when it is validated against input parameters. The architecture of the
model for fault intensity detection is shown in Figure (1).

Figure 1. Architecture of the proposed model for fault detection of complex systems using neurofuzzy controller

The transformation of raw data to fuzzy value using membership functions marked the beginning of
fuzzification. The main issue is to generalize the parameters of membership functions for fuzzification from the
training examples to the entire domain by training the neural network.
Let FNN denote the function directly encoding the network operation. The network is trained so that FNN
approximates f,
FNN ≈ f (1)
where f is the function that gives the most optimal alarm values. For the purpose of training, a set of training
examples are selected from the domain. Each example is represented by (x,y) where y=f(x). The ideal goal is
that after training,
y = f(x) = FNN (x) (2)
for every x in the domain, not limited to the training set. Sum of squared error is most commonly used to
calculate accuracy of FNN .

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 209

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

(3)
SSE can be a high value because of mainly the following four cases of errors in parameters of membership
functions (µ1 and µ2) of fuzzification process.

Figure 2. Four possibilities of errors

(1) Case A- The desired value is located to the left of µ1, but the actual output lies between µ1 and µ2.
(2) Case B - The desired value is located between µ1 and µ2, but the actual output lies to the left of µ1.
(3) Case C - The desired value is located between µ1 and µ2, but the actual output lies to the right of µ2.
(4) Case D - The desired value is located to the right of µ2, but the actual output lies between µ1 and µ2.
In all the cases the output value is being produced by incorrect fuzzy sets, and hence the shape of input
membership functions needs to be adjusted. The design of the network architecture and the training algorithm is
aimed at improving this capability of reducing the squared error.
The architecture of the proposed model for training can be compared with the architecture of a neural network
where each layer of the neural network corresponds to each step of designing a fuzzy inference system. The
architecture consists of input layer, which corresponds to receiving the actual crisp value input. The next hidden
layer is the fuzzification layer that uses the input membership functions to convert the crisp inputs to fuzzy
inputs. This layer is followed by the layer with neurons Ri, which are the rules or the constraints of the complex
system. The next layer is the one with the function of defining the output membership function. The output layer
uses defuzzification method to convert all the fuzzy outputs obtained to one crisp value.

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 210

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

Figure 3. Training the fuzzy controller by back-propagating the error to the neural-like architecture.

The X-, R-, and C-modules are the neurons in a layered neural network and the µ - and v-values as the adaptable
weights of the network. This corresponds to the parameters passed to the membership functions during the
fuzzification stage and the defuzzification stage of designing a fuzzy inference system. The X-module layer is
the input layer of a multi-input neural network whereas the C-module layer is the output layer. The R-module
layer serves as the hidden or intermediate layer that defines the fuzzy rule base. The fact that one µ -module can
be connected to more than one R-module is equivalent to the connections in a neural network. This is of key
importance for keeping the structural integrity of the fuzzy controller intact.
For training a neural network, back-propagation algorithm is used. For each epoch, error value is calculated as
per Equation (3). If the error value is less than threshold, the parameters of membership functions are fixed. If
the error value is greater than threshold, then parameters are changed and the training continues, by distributing
back the control value among the control rules. The training continues until the error becomes less than
threshold or until the desired number of epochs are reached. The main application of neural network is for the
fine tuning of the parameters of the membership functions as a reverse mechanism deduced from the forwarding
inference machinery of the fuzzy logic controller.
The selection of parameters of the membership functions of fuzzification governs the rules to be triggered. An
inappropriate parameters fixing of membership functions result in wrong selection of the rules to be triggered
which in turn results in the generation of a wrong crisp output. This deteriorates the performance of the model
when the model finds applications in areas like fault tolerance. Hence in this work, the accuracy of the
parameters are judged and fixed based on the calculated SSE.

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 211

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

Algorithm 1: for training the fuzzy control system using neural network.

Input : .DST, MFN ,MFT,µ1, µ2, ET, NE

Output : µ1’, µ2’

Step 1 : Define the number of membership functions(MFN) and type of membership functions (MFT) for
fuzzification of input variables.
Step 2 : Define the initial parameters for membership functions. (µ1 and µ2)
Step 3 : Define the training data set.(DST)
Step 4 : Set the error tolerance value or the threshold value.(ET)
Step 6 : Define the number of epochs(NE).
Step 7 : # Epochs =1.
Step 8 : Compute EA = SSE. based on Equation (3).
Step 9 : If the error EA > ET go to step 10, otherwise go to step 13.
Step 10 : #Epochs = # Epochs + 1.
Step 11 : Change the parameters of the membership functions by training the system using back-propagation
algorithm .
Step 12 : Repeat Step 8 – Step 11 until
(EA < ET or # Epochs > NE or EA becomes a constant)
Step 13 : Take the paramaters of µ1 and µ2 as the new parameters µ1’ and µ2’ for defining the fuzzy
membership function.

A fuzzy inference system can be designed with the new parameters µ1’ and µ2’ for the membership function of
the fuzzification process of the input variables. The fuzzified values of the crisp inputs trigger one or more rules
from the fuzzy rule base of the inference system. All the rules of the fuzzy rule base are assigned weights based
on the priority of corresponding constraints. This model with fuzzy rules and membership functions identified
using neural network can be used for fault detection when the model is validated against instance values, with
great accuracy.
IV. Simulation results for fault detection in a thermal power plant system
The system taken for study is a coal-fired thermal power plant. The primary fuel fed to the power plant is coal,
which is used to generate electricity. The thermal power plant is a complex system and it is constrained with a
set of parameters. Coal is fed to the pulverizer component and is pulverized to fine granules. This is given to the
boiler component. The steam generated from the boiler rotates a turbine to which a generator is attached. This
generates electricity.
Literature reveals that many mathematical coal mill modeling techniques have been proposed and implemented
which resulted in optimized milling process[18]. But in this work, an attempt is made to optimise an object
oriented model of a coal mill model by using an optimal fuzzy logic controller to represent the fuzzy constraints
of the system. The proposed model has the advantage of simulation and testing of the model before actually
building the system.
In this work, the subsystem taken for study is “Boiler”, and the three dependent parameters of Boiler are
“Steam_Inlet_temperature”, “Steam_Inlet_Pressure” and “Feed_Water_Temperature”. These parameters are
constrained to a set of domain values. In a fuzzy controlled system, faults of different intensities can be
identified based on the values of these parameters. Table (1) shows the permissible range for these parameters.
Table 1 . Domain range of three input parameters for Boiler

Input parameters Minimum Maximum

Steam_Inlet_temperature 320 340
Steam_Inlet_Pressure 20 30
Feed_Water_Temperature 235 255

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 212

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

The crisp value of these parameters fed to the model needs to be validated against the constraints defined as
fuzzy rules. The rule selection is on the basis of the parameters of membership functions. This ultimately
governs the intensity of faults detected. Hence choice of parameters for membership function is very crucial.
In this paper, neural network training algorithm of back-propagation is used to identify the most optimal
parameters for input and output membership functions. For accurate parameter identification, the control
system is trained with training inputs as explained in the Algorithm1.
Triangular membership functions shown in Equation (4) gives the best drive performance while implementing
fuzzy logic in control systems.[19].Hence, the input variable of Steam_Inlet_temperature, Steam_Inlet_Pressure
and Feed_Water_Temperature are defined with three triangular membership functions.

(4)
The linguistic variables associated with each input parameter are LO(Low), RT(Right) and HI(High). Table (2)
shows the parameter range for the linguistic variables associated with each input.
Table 2 . Parameter range of input linguistic variables

Ling.Var Parameters
St_Inlet_Temp St_Inlet_Pr Feed_Water_Temp
LO [320 330] [20 24] [235 245]
RT [325 335] [23 27] [240 250]
HI [330 340] [26 30] [245 255]
Training data set with expected alarm values obtained from thermal power plant site is used to train the model
for many epochs. Increase in the number of epochs increases the accuracy of the model as the average training
error decreases. It is seen that as the number of epochs increases, the training error decreases and remains
unchanged after a specific epoch.
Table 3. Sample training dataset

No. St_Inlet_Temp St_Inlet_Pr Feed_Water_Temp Alarm

1 320 20 235 -1.5
2 330 26 246 0
3 340 30 255 1.5
4 336 28 252 1.3
5 334 26.5 250 1
6 332 27 248 0
7 325 22 240 -1.3
8 338 28 250 1.2
9 335 27 248 0.98
10 329 25 244 -0.98
The output linguistic variable “alarm” and membership function are defined as shown in Table(4) and Figure(4)
respectively.

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 213

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

Figure 4. Output membership function graph

Table 4. Range of output linguistic variable “alarm”

Alarm(output variable)
NEG -1.5 -0.5
LN -1 0
ZR -0.5 0.5
LP 0 1
POS 0.5 1.5

Fuzzy rules are defined with the three parameters as antecedents and alarm as the consequent, shown in Tables
5a, 5b and 5c.
Table 5a. Fuzzy rules when St_Inlet_Temp=LO

St_Inlet_Temp=LO
F-W-temp
St_Inlet_Pr
LO RT HI
LO NEG LN LN
RT LN LN LN
HI LN LP LP

Table 5b. Fuzzy rules when St_Inlet_Temp=HI

St_Inlet_Temp=HI
St_Inlet_Pr F-W-temp
LO RT HI
LO LN LP LP
RT LP LP LP
HI LP LP POS

Table 5c.Fuzzy rules when St_Inlet_Temp=RT

St_Inlet_Temp=RT
St_Inlet_Pr F-W-temp
LO RT HI
LO LN LN LP
RT LN ZR LP
HI LP LP LP

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 214

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

The defuzzification uses the weighted average method given in Equation(5) and the results are shown in Table
(6).

(5)
Table 6.Comparison of expected and obtained alarm values

St_Inlet_Temp St_Inlet_Pr F-W-Temp Alarm (Fuzzy Alarm (Expected

controller) x Result) y
326 23 242 -1 -1.3
330 25 245 0 0
336 29 252 1 1.3
330 26 246 0.146 0
336 28 252 1 1.3
334 26.5 250 0.751 1

The sum of squared obtained is 0.3533, which is not a negligible value.

Figure 5. Graph showing comparison of expected and obtained outputs.

The high value of SSE reveals that the membership parameters are not optimal. Hence the FIS is trained using
the back-propagation algorithm of neural network. Figure(6) shows the architecture of training FIS of Boiler
with three parameters and twenty seven rules.

Figure 6.Architecture for training the FIS of Boiler with three parameters

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 215

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

Table 7 . Training error through epochs

Epochs MSE
10 1.055
20 0.999
50 0.837
100 0.599
200 0.298
400 0.142
500 0.107
MSE decreases with increase in epchs and the corresnding values is given in Table 7. A graphical reresentation
is shown in Figure(7).

Figure 7. A graphical representation of decrease in error rate with epochs

Table 8. Comparison of expected and obtained alarm values

St_Inlet_Temp St_Inlet_Pr F-W-Temp Alarm (Trained Alarm (Expected

Fuzzy controller) x Result) y
320 20 235 -1.5 -1.5
330 26 246 -0.037 0
340 30 255 1.5 1.5
336 28 252 1.22 1.3
334 26.5 250 1.06 1
332 27 248 0.078 0
325 22 240 -1.31 -1.3
338 28 250 1.27 1.2
335 27 248 0.857 0.98
329 25 244 -0.958 -0.98
The SSE value now reduced to 0.038, which is a negligible value.

Figure 8: Graph showing comparison of expected and obtained outputs.

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 216

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

For a constrained complex system, the constraints or the rules may be of different priority levels. Here, the
priority of constraints is mapped to the weights of fuzzy rules in the fuzzy rule base. The priority p of rule i, p(i)
is defined based on Equation(6).

(6)
The defuzzification is based on equation(7), which is a modified weighted average method with priority p(i).

(7)
Table 9. Comparison of expected, Trained FIS and prioritised trained FIS output values

St_Inlet_ St_Inlet_Pr F-W-Temp Alarm(Trained Alarm (Expected Alarm (Trained

Temp Fuzzy controller) Result) y Prioritised fuzzy
x controller)
320 20 235 -1.5 -1.5 -1.5
330 26 246 -0.037 0 -0.004
340 30 255 1.5 1.5 1.5
336 28 252 1.22 1.3 1.3
334 26.5 250 1.06 1 1.01
332 27 248 0.078 0 0.04
325 22 240 -1.31 -1.3 -1.33
338 28 250 1.27 1.2 1.21
335 27 248 0.857 0.98 1.01
329 25 244 -0.958 -0.98 -0.98

Figure 9: Graph showing comparison of expected, trained and weighted(prioritised) outputs.

The SSE value on trained prioritized fuzzy controller, gave 0.0037, which is a negligible value. Hence, the
parameters of membership functions for each of the input linguistic variable are considered the optimal. The
above results reveal that a more accurate alarm value is obtained from the fuzzy control system with weighted
fuzzy rules.
V. Conclusion
A fuzzy controller is designed and implemented for fault detection of Boiler subsystem of a coal-fired thermal
power plant system. A rule based system can be used for fault detection if the rules are crisp and precise in
nature. We have proposed and implemented an intelligent fuzzy controller for handling imprecise and uncertain
situations and proved its suitability for fault detection of a constrained complex system. The unknown dynamics
of fuzzy rules are identified using the soft computing technique of neural networks. The priority of constraints
of the complex systems defines the weights of the fuzzy rules in the rule base. The results reveal that the

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 217

 Priya R et al. / Indian Journal of Computer Science and Engineering (IJCSE)

prioritised and trained fuzzy control system can be used for generating alarms of different intensities from -1.5
to 1.5, based on which the intensity of the fault can be identified. This control system can replace the existing
rule based system which generates only one specific alarm for any constraint violated. The control result can be
improved by resizing the fuzzy sets and finer tuning for the membership functions.
References
[1] Lee, Ching-Hung, and Ching-Cheng Teng. "Identification and control of dynamic systems using recurrent fuzzy neural networks."
IEEE Transactions on fuzzy systems 8.4 (2000): 349-366.
[2] Rattan, Kuldip S., T. Brehm, and Gurpreet S. Sandhu. "Analysis and Design of a Proportional Fuzzy Logic Controller." Proceedings of
the International Fuzzy Systems and Intelligent Control Conference. 1994.
[3] Vieira, Jose, F. Morgado Dias, and Alexandre Mota. "Neuro-fuzzy systems: a survey." 5th WSEAS NNA International Conference.
2004.
[4] Sharma, Disha. "Designing and modeling fuzzy control Systems." International Journal of Computer Applications 16.1 (2011): 46-53.
[5] Lee, Minho, Soo-Young Lee, and Cheol Hoom Park. "Neuro-fuzzy identifiers and controllers." Journal of Intelligent & Fuzzy Systems
2.1 (1994): 1-14.
[6] W. J. M. Kickert and H. R. Van Nauta Lemke, “Application of a fuzzy logic controller in a warm water plant,” Automatica, vol. 12,
pp.301–308, (1976).
[7] Snehal N. Rajguru, Pritesh Shah, NeelaRayavarapu and PriyankaTupe, “Application of fuzzy controllers for flow control processes in
chemical industries” Journal of Chemical and Pharmaceutical Research, 6(12):pp.403-10 Research Article ISSN : 0975-7384
CODEN(USA) : JCPRC5, (2014).
[8] I. Baturone, F. J. Moreno-Velo, S. Sanchez-Solano, and A. Ollero, “Automatic design of fuzzy controllers for car-like autonomous
robots,” IEEE Trans. Fuzzy Syst., vol. 12, no. 4, pp. 447–465, (Aug. 2004).
[9] R. J. Kornblum and M. Tribus, “The use of Bayesian inference in the design of an endpoint control system for the basic oxygen
steelfurnace,” IEEE Trans. Syst., Man, Cybern., vol. SMC-6, no. 2, pp. 339–348, (Mar./Apr. 1970).
[10] R. M. Tong, M. B. Beck, and A. Latten, “Fuzzy control of the activated sludge wastewater treatment process,” Automatica, vol. 6, pp.
695–701,(1980).
[11] Om Prakash Verma and Himanshu Gupta, “Fuzzy Logic Based Water Bath Temperature Control System” Volume 2, Issue 4,
International Journal of Advanced Research in Computer Science and Software Engineering Research Paper, (April 2012).
[12] Isizoh A. N., Okide S. O, Anazia A.E.. Ogu C.D. “Temperature Control System Using Fuzzy Logic Technique”.(IJARAI)
International Journal of Advanced Research in Artificial Intelligence, Vol. 1, no. 3, (2012).
[13] Snejana Yordanova, Daniel Merazchiev, and Lakhmi Jain “A Two-Variable Fuzzy Control Design With Application to an Air-
Conditioning System” IEEE transactions on fuzzy systems, Vol. 23, no. 2, April 2015.
[14] Kapil Dev Sharma, M. Ayyub, Sumit Saroha, Ahmad Faras “Advanced Controllers Using Fuzzy Logic Controller (FLC) for
Performance Improvement” International Electrical Engineering Journal (IEEJ) Vol. 5 no.6, pp. 1452-58,(2014).
[15] T. Haruki and K. Kikuchi, “Video camera system using fuzzy logic,” IEEE Trans. Consumer Electron., vol. 38, no. 3, pp. 624–634,
(Aug.1992).
[16] Y. M. Lee, S. I. Jang, K. W. Chung, D. Y. Lee, W. C. Kim, and C. W.Lee, “A fuzzy-control processor for automatic focusing,” IEEE
Trans.Consumer Electron., vol. 40, no. 2, pp. 138–144,(May 1994).
[17] Jyh-Shing Roger Jang. Input Selection for ANFIS Learning, Proc.5th IEEE International Conference on Fuzzy Systems. 1493-
1499.(1996).
[18] Wei, Jianlin, Jihong Wang, and Shen Guo. "Mathematic modeling and condition monitoring of power station tube-ball mill systems."
2009 American Control Conference. IEEE, 2009.
[19] Zhao, Jin, and Bimal K. Bose. "Evaluation of membership functions for fuzzy logic controlled induction motor drive." IECON 02
[Industrial Electronics Society, IEEE 2002 28th Annual Conference of the]. Vol. 1. IEEE, 2002.
[20] Snejana Yordanova, Daniel Merazchiev, and Lakhmi Jain, “A Two-Variable Fuzzy Control Design With Application to an Air-
Conditioning System”, IEEE Transactions on fuzzy systems, Vol. 23, no. 2, (April 2015).

ISSN : 0976-5166 Vol. 7 No. 5 Oct-Nov 2016 218