Neural networks and fuzzy logic in electrical engineering control courses

Abstract
Control system education must include experimental exercises that complement the theory
presented in lectures. These exercises include modelling, analysis and design of a control system. Key
concepts and techniques in the area of intelligent systems and control have been discovered and
developed over the past few decades. While some of these methods have significant benefits to offer,
engineers are often reluctant to utilise new intelligent control techniques, for several reasons. In this
paper fuzzy logic controllers have been developed using speed and mechanical power deviations, and
a neural network has been designed to tune the gains of the fuzzy logic controllers. Student feedback
indicates that theoretical developments in lectures on control systems were only appreciated after the
laboratory exercises.
Keywords
fuzzy logic; intelligent control; modelling; neural networks
Undergraduate students in computer science learn best when they are given the
opportunity to apply software concepts to real world systems, and intelligent control
applications present attractive possibilities for giving them such opportunity. An
example of how to take advantage of these possibilities is given in this paper, which
describes a specific neural network technique that has been developed and applied
to the problem of tuning fuzzy controllers.
To master the technique, the students start by learning about neurons and fuzzy
logic, but they soon find themselves ‘training’ a multilayered neural network that
they themselves have built.
The electronic computer has been used extensively in the electric power industry
from the moment the computer became commercially available. This was a natural
development since the size and complexity of most power system problems make
the computer an essential tool for the electric power system analyst and designer.
Electric power system analysis tools, with computational power similar to or
better than programs first developed during the 1960s and 1970s, and easy-to-use,
attractive, and providing a graphical user interface, are just becoming available.1,2
The main difficulty in a control course is the large amount of mathematical models
and equations. The repetitive algebra and the necessity for a complete understanding
of the physical concepts embedded in the equations require the use of a suitable
computational tool. The introduction of MATLAB in this course is due to its facility
to build up mathematical functions and also due to its powerful graphical user
interface in order to display the results.
We describe a laboratory exercise required in a control class and the relationship
between the exercise and the theory presented in lectures. We conclude with a
summary of our experiences with the laboratory exercises and the feedback received
from our students and graduates.
Intelligent control
Major concepts and techniques in the area of intelligent systems and control have
been discovered and developed over the past few decades.3 While some of these
methods have significant benefits to offer, engineers are often reluctant to utilise new
intelligent control techniques for several reasons:
l there has been a lack of rigorous engineering analysis to verify, for example,
stability properties and performance characteristics;
2 there is not an established track record for the reliability and robustness of such
techniques;
3 there has not been comparative analysis to determine their advantages/
disadvantages relative to conventional methods; and
4 the approaches are not widely understood by practising engineers. The relative
lack of attention given to the potential of intelligent control is cause for some
concern, indicating a definite need for applications-directed research and education
in these areas.
Curricula for control engineering programs have undergone substantial change in
the past years as modern techniques for analysis and design find their way into these
courses. lt is quite natural then, that newer technologies such as intelligent control
should be introduced into university curricula. Along with the continuously evolving
curricula, there remains a constant in control engineering education: the recognised
need for laboratory experience in the curricula. More and more examples of
high-quality control laboratories are appearing in universities worldwide.4 Moreover,
more and more educators recognise the importance of a complete educational experience
involving theory and practice.
With these thoughts in mind, it has been our goal to bring the newest technologies
into the curricula, through both lecture and laboratory courses. With regard to
the treatment of intelligent control in the courses our intent has not been to give an
in-depth treatise on the theory of fuzzy sets and neural nets. We found that electrical
engineering undergraduates have little difficulty in coming up to speed in the
area in a relatively short amount of time.
The graduate students are exposed to the more advanced topics, and at a higher
level of sophistication. Several projects in simulation and design analysis are given.
The graduate students are required to complete the entire lecture-laboratory course.
Laboratory exercises
We believe that effective control system education must include experimental exercises
that complement the theory presented in lectures. Preferably, the exercises
should include the design and implementation of a control system. Limited resources
make it impossible to offer a laboratory course with every control class in most electrical
engineering curricula. The students are divided into groups and each group
must independently complete the design project and submit a formal report summarising
their results and experiences. Each student must submit an individual
2 F. Jurado, A. Caño and M. Ortega
International Journal of Electrical Engineering Education 40/1
commentary on the exercise and the experimental results obtained. Student feedback
indicates an increased appreciation of the lecture material and an awareness of the
limitations of the theory and simulation that was lacking prior to the introduction of
laboratory exercises.
The classes include highly mathematical theoretical derivations which must be
justified to our students. The number of credit hours allocated to control is already
high relative to other areas and there is no justification for increasing it.
Nevertheless, the courses include highly mathematical developments which many
seniors find challenging. The courses include elements of computer-aided control
system design. However, some of the limitations of the theory and design recipes
are not easily appreciated from the simple problems that can be addressed in lectures
and exams.
The exercises include modelling, analysis and design of a control system. Practising
engineers often obtain mathematical models of physical systems by evaluating
the system parameters in the laboratory, design a controller based on the
mathematical model, then use a more complex model in extensive simulations. The
simulations provide an economical means of testing the system before implementation.
Finally, some modification of the design may be necessary to obtain acceptable
performance. Trial and error may be necessary at each stage before the
desired results are obtained. Our goal was to develop a take-home exam including
laboratory exercises that mimic this general sequence of events. Design problems
suitable for in-class exam must involve little or no trial and error due to time
limitations.
Using a tool such as MATLAB, the students then design a controller for the gas
motor based on its model. They learn that the model used for design can be much
simpler than that used for simulation provided that the performance of the closedloop
system is checked using the more complex simulation model. The students are
given guidelines for the selection of the design specifications but the design specifications
themselves are not provided. An important part of the exercise is that the
students have to experiment with the system to be able to choose realistic design
specifications. In practice, the design engineer may have some freedom in choosing
the specifications so that the design criteria lie in acceptable ranges. In addition, the
acceptable ranges for the design criteria are based on an understanding of the physical
system and its normal operating conditions.
Gas motor control
Natural gas as a fuel for diesel engines offers the advantage of reduced emissions
while retaining the high efficiency of the conventional diesel engine. The engine can
operate at high compression ratio with a wide range of gas composition. A disadvantage
of natural gas use in diesels is the high auto-ignition temperature. Thus,
ignition assistance is needed. Fuel rates are always expressed in terms of higher
heating value. Heat balance accounts for all the thermal energy involved in the
process of converting fuel to energy. This energy is then converted into mechanical
work, by expanding the gas through the motor. The majority of medium- and slowspeed
engines are turbocharged, using axial-flow or mixed-flow designs. The data
Neural networks and fuzzy logic 3
International Journal of Electrical Engineering Education 40/1
required for the model consist of a set of algebraic functions which relate the significant
engine variables. These functions are obtained from engine test data.5–7
The gas motor controller is modelled by a set of simultaneous linear differential
equations relating the engine speed, mechanical power, reference speed and load
sharing signal with the fuel demand signal.8–10
The gas motor controller regulates both the gas motor and the gas motor generator.
For the purpose of this exercise only modulating control of the mechanical side
of the gas motor is of interest. The simplified model of the gas motor controller in
this exercise consists of two inputs and one output. Inputs to the controller, which
are outputs from the gas motor model, are the mechanical power delivered by the
motor Pmec and the rotation speed of the gas motor w, related to the electrical frequency
of the generator. The output from the controller is the fuel demand signal Fd.
The block diagram of the gas motor control system is presented in Fig. 1. The
diagram consists of two feedback controllers. LVG stands for Least Value Gate,
which transmits the minimum of two incoming signals.
PID controller
Despite huge advances in the field of control systems engineering, PID still remains
the most common control algorithm in industrial use today. It is widely used because
of its versatility, high reliability and ease of operation.11Astandard method of setting
the parameters is through the use of Ziegler-Nichols’ tuning rules.12 These techniques
were developed empirically through the simulation of a large number of process
systems to provide a simple rule. The methods operate particularly well for simple
systems and those which exhibit a clearly dominant pole-pair, but for more complex
4 F. Jurado, A. Caño and M. Ortega
International Journal of Electrical Engineering Education 40/1
Fig. 1 Gas motor block diagram.
systems the PID gains may be strongly coupled in a less predictable way. For these
systems, adequate performance is often only achieved through manual and heuristic
parameter variation.
Neuro-fuzzy logic controller
Unlike the classical control design, which requires a plant model for designing the
controller, fuzzy logic incorporates an alternative way which allows one to design
a controller using a higher level of abstraction without knowing the plant model.
This makes the fuzzy logic controller (FC) very attractive for ill-defined systems or
systems with uncertain parameters.13
The recent growth in attention to neural networks (NNs) has led to many suggestions
for combined use of fuzzy logic and neural networks in intelligent
control.14,15 In a multi-layer neural network of the feed-forward type, input nodes
record the features and pass activation values to the output layer through a hidden
layer. The addition of the hidden layers to the two layer perceptron networks allows
these networks to represent any continuous mapping from input to output.16,17 An
appropriate training technique adjusts the connection weights of the network to
improve the match between the output of the network and the correct results.
To design the FC some variables, which can represent the dynamic performance
of the system, should be chosen to be fed as the inputs. In addition to the proper
input signals, signal gains and fuzzy subsets should be defined. It is common to use
the output error and the rate of derivative of the output as controller inputs.18,19 In
this paper, the motor speed deviation (Dw) and its derivative (Dw¢), the acceleration,
are considered as the inputs of the first FC and the mechanical power deviation
delivered by the motor (DPm) and its derivative (DPm¢) as the inputs of the second
FC.
Subsequently, Dw, Dw¢, DPm and DPm¢ signals pass through four appropriate gains
or scaling factors, and then are fed to the FCs. The outputs of the controllers are
also scaled by passing through the output gains. To convert the measured input variables
of the FCs into suitable linguistic variables, seven fuzzy subsets are chosen.
Membership functions of these subsets are bell-shaped. Figure 2 shows the membership
functions. In this paper, both inputs of the FCs have seven subsets. Thus,
two fuzzy rule tables with forty-nine rules are constructed. Figure 3 illustrates the
control surface. The centre of gravity method is employed.
Sequential de-centralised control means design of each modulation controller one
after the other, so that dynamics of previously designed controllers are taken into
account in designing the next controller.20
Tuning the fuzzy controllers
In order to tune the FCs, the Dw is scaled according to the following relation:
Dw* = GewDw and Dw¢, Dw¢* = GrwDw¢. The DPm is scaled according to the relation:
DPm* = GePmDPm and DP¢m, DP¢m* = GrPmDPm. Also, the output of the first FC is scaled
by Guw and the output of the second FC is scaled by GuPm. In the aforementioned
relations, Gew, Grw,Guw, GePm,GrPm and GuPm are the scaling factors or gains.
The gains of the FCs are tuned with a neural network, making the FCs adaptable