41
CHAPTER 4
LINEAR AND NON-LINEAR CONTROLLER
A Quasi-Resonant Buck Converter experiences a high degree of
non-linearity in its control characteristics mainly due to switching action or
fluctuations in system parameters and to achieve the desired regulation
irrespective of line and load disturbances feedback control is indispensably
required. A common feature of control methodology is that the algorithm is
analytically described by equations such as algebraic, difference, differential
and so on and the synthesis of such algorithm requires a mathematical model.
Hence, it is the responsibility of an engineer to determine how to accurately
describe a system mathematically and how to formulate proper assumptions
and approximations so that the system shall be realistically characterized by a
mathematical model. If the assumptions pertaining to the mathematical model
are satisfied, many of the model-based control techniques such as Bode or
Nyquist plot provide good stability, robustness to model uncertainties and
disturbances, and speed of response. In real-life applications or in a complex
process such as cement plant, nuclear reactor and the like requires
complicated mathematical analysis, floating point algorithm and complex
branching which makes the implementation of mathematical model very
tedious and time consuming.
A linear controller is probable to be unstable if the required range
of operation is high because the non-linearities in the system cannot be
properly compensated. In addition, if the model is ill-defined or if the model
has parameters whose value is partially identified, the algorithm based on
such incomplete information will not give satisfactory results and the linear
42
controller exhibits degradation in its performance or instability. This reality,
inevitably promotes the endeavour to control methods that will more or less
incorporate the non-linear dynamics into the design.
A non-linear controller handles the nonlinearities in large range of operation
and such nonlinearities shall be intentionally introduced into the controller
part of a control system so that the model uncertainties may be tolerated.
Advancement in computer technology made the implementation of non-linear
control methodologies such as adaptive control (Gopal 2009), sliding mode
control (Ahmed et al 2003), (Lai et al 2005), current programmed control
(Jingquan Chen et al 2003), (Erickson and Maksimovic 2006), intelligent
control such as Fuzzy and Neuro-Fuzzy control, etc., a relatively simple task.
4.1 PI CONTROL
A PI or Proportional-Integral controller is a linear controller which
fuses the properties of both P and I controller and the algorithm provides a
balance of complexity and capability in order to be widely used in industrial
applications pertaining to process control. The proportional value determines
the reaction to the recent error whereas the integral value determines the
reaction based on sum of recent error and the weighted sum of two actions is
used to adjust the process through a control element.
The Equation (4.1) that describes the nature of P controller is
u(t) = KP e(t) (4.1)
where Kp is the proportional gain, e(t) is the error and u(t) is the perturbation
in output signal of controller from the base value corresponding to normal
operating conditions; the base value is adjusted to produce zero error under
the condition of no disturbance and changes in set-point. It with no integration
property always exhibit steady-state error in the presence of disturbances and
changes in set-point. The error, of course, must be made negligibly small by
43
increasing the gain of proportional controller; however, as the proportional
gain is increased, the closed-loop performance of the system shows a
relatively maximum overshoot and long settling time. To remove steady-state
offset in controlled variable of a process, an extra intelligence is added to the
P controller and it is called the integral action which induces robust voltage
regulation at output against parameters uncertainties and external disturbances
(Ramirez et al, 2001), (Zafiriou and Morari 1989). The controller is a PI
controller whose mathematical notation is depicted in Equation (4.2).
ª º
K i ³0
u(t)= K p «e(t) + »
t
1
¬ ¼
e(t) dt (4.2)
where K i is the integral or reset time. A linear PI controller has thus two
tuning parameters namely Kp and Ki; the integral action (Gopal 2009) in this
controller removes steady-state offset in the controlled variable of a process.
4.2 INTELLIGENT CONTROL
It is necessary to increase sophistication of the controller to handle
the complexity of non-linearities such as backlash, coulomb friction, dead
zone or saturation in the system and in order to extend the operating range of
the controller to wide variations in either line or load, (Criscione et al 2001) it
is more appropriate to use a methodology from Artificial Intelligence so as to
achieve intelligent control action. Intelligent control such as Fuzzy or multi-
valued logic was propounded by Dr. L.A. Zadeh, Professor, University of
California, Berkley, California 94720 (Bose 1994), (Hua Li and Madan Gupta
1995), (Buckley 2011), (Yen and Langari 1999), (Zimmerman 2011) and is
largely rule-based because the dependency involved in its deployment is too
complex to permit an analytical representation. To deal with such
dependency, the mathematics of fuzzy system integrates the experience and
knowledge gained in the operation of a similar plant into control algorithm
44
and is of great value for problems where mathematical model of system is
difficult to be obtained due to complexity, non-linearity and imprecision
(Perry et al 2005), (Feng et al 2002); however, the expense lies in high
computational density (Weidong Xiao and Dunford 2004). The algorithm
consists of a set of IF-THEN rules (Fumio Ueno et al 1991), (Mendel 1995)
and is based on expert’s knowledge and is normally of the form:
IF (process state) THEN (control action) or otherwise
IF (a set of conditions are satisfied) THEN (a set of consequences
can be inferred)
The process state part of rule is called the antecedent part and it
contains a description of the process state at the kth sampling instant whereas
the control action part of the rule is the consequent part (Berenji and Khedkar
1992) and it contains a description of the control variable produced as per the
particular process state. The collection of such fuzzy rules which are
expressed as conditional statements forms the rule base. There is no procedure
for deciding the optimal number of fuzzy rules since a number of factors is
involved in the decision such as performance of the controller, efficiency of
computation, human operator’s behavior, and the choice of linguistic
variables. In practice, it is important to ensure the consistency of fuzzy rules
in order to minimize the possibility of contradiction. For instance, in the case
of two-input-single-output fuzzy system, the fuzzy control rule is of the form:
R1: IF x is A1 AND y is B1 THEN z is C1
R2: IF x is A2 AND y is B2 THEN z is C2
R3: IF x is A3 AND y is B3 THEN z is C3
R4: IF x is A4 AND y is B4 THEN z is C4
45
and so on; this fuzzy rule- set may be combined as one single rule by means
of an union operator and shall be expressed as R= R1 U R2 U……..U RN in
which the fuzzy relation R is regarded as a fuzzy relation whose function is to
map elements from one universe, say U, to another universe, say V through
the Cartesian product of two universes and its membership function is
depicted in Equation (4.3) (Fang Hsien Wang and Lee 1995) as
ȝ R (x,y) = ȝ AXB (x,y) = min{ȝ A (x,y),ȝ B (x,y)}; x U,y V (4.3)
Here, x, y and z are the linguistic variables (Zadeh 1973)
representing two process state variables and one control variable respectively
in natural language; the proper choice of variables is essential to the
characterization of the operation of a fuzzy system. The variables A1, B1 and
C1 are the linguistic values of the linguistic variables x, y and z in the
Universe of discourse U, V and W respectively. A Fuzzy Controller
(Drainkov et al 1997) incorporates attractive features such as good
performance and simplicity and it does not require small signal model of the
converter and is easier to implement (Smyej and Cheriti 1999); the
performance of such controller depends on rule size and tuning parameters
(Sung-Hoe Huh and Gwi-Tae Park 1999). In addition, the control algorithm is
generalized and hence same control rules may be applied to various DC-DC
converters with minor modifications depending upon the topology of the
converter and parameters (Arulselvi et al 2006). It is highly adaptive in nature
(Lin and Hua 1993) and can also exhibit increased reliability, robustness in
the face of changing parameters, saturation effects and external disturbances
and so on. It extends the control capability even to those operating conditions
such as large signal dynamics and parameter variations where linear controller
fail (Corcau et al 2010); fundamentally, there are two fuzzy models namely
[Link] model (Takagi and Sugeno 1985) and [Link]
model in which the former model possess an excellent capability in complex
and uncertain system description and is suitable for modeling the nonlinear
system by fuzzy local models combined using an inference mechanism
46
corresponding to various operating points (Yu and Sun 2001); it has found
applications in control, prediction and inference (Jang 1993) such as
information retrieval, decision-making, database management, signal
processing and so on (Jang and Sun 1995).
Figure 4.1 clearly depicts the basic configuration of
[Link]’s Fuzzy controller (Idiarie et al 2004) whereas Figure 4.2
illustrates the architecture of a Fuzzy control system (Abdelnour, 1991). It
consists of four components (Chuen-Chien Lee 1990), (Corcau and
Constantinache 2007), (Nik Ismail et al 2010) namely a fuzzifier, a
knowledge base containing rule and data base (Melin and Vidolov 2003), a
decision-making logic and a defuzzifier; in brief, fuzzifier and de-fuzzifier are
used to communicate with the real world; a defined set of rules is used for
governing the converter in order to obtain the desired performance at the
output and the inference engine is responsible for making fuzzy decisions
based on the given input and a set of rules (Ramos et al 2000); the operation
of Fuzzy controller does not rely on how accurate the mathematical model is,
however on how effective the rules are or otherwise, the operation of Fuzzy
controller is based on expert’s knowledge of the plant instead of a precise
mathematical model (Passino and Yurkovich 1997). It thus generated a good
deal of interest in applications relating to power electronics and electrical
drive systems due to its capability of fast computation with high precision.
Figure 4.1 Basic configuration of [Link]’s Fuzzy controller
47
Figure 4.2 Architecture of Fuzzy control system
48
In spite of the usefulness of Fuzzy control, standard method doesn’t
exists for transforming human knowledge into the rule base and database of a
fuzzy inference system; in addition, stability analysis of a fuzzy system is not
easy and parameter tuning is generally a time-consuming procedure due to the
non-linear and multi-parametric nature of fuzzy systems (Ho Jae Lee et al
2001). Hence, the neural network which has a large number of interconnected
processing nodes to demonstrate the ability to learn and generalize from
training data and which has many successful applications such as [Link]
and [Link]’s NETtalk for text- to speech conversion; [Link],
[Link] and [Link]’s Neocognitron’s for visually recognizing hand-written
Arabic numbers (Fukushima et al 1983); [Link]’s vision processing
network to synthesize coherently three-dimensional form, color and
brightness perception; [Link], [Link] and [Link]’s control of
cart-pole problem (Barto et al 1983); [Link]’s control of an inverted
pendulum and [Link]’s non-linear self-tuning adaptive control (Chen
1989) is integrated with fuzzy logic in order to provide a promising
methodology known as Neuro-Fuzzy Inference System.
A Neuro-Fuzzy Inference System is a multi-layered connectionist
network (Lin and George Lee 1991) which combines parallel computation
and learning abilities of neural networks (Tang et al 2011), (Narendra and
Annasamy 1989) with the human-like knowledge representation and
explanation ability of fuzzy logic control and decision systems (Cheng-Jian
Lin et al 2008) and the structure of such system has input terminals, hidden
layers to represent membership functions and fuzzy rules and an output layer.
It is isomorphic to fuzzy logic control systems in terms of their functions and
is also referred to as universal approximators (Wang and George Lee 2002).
In general, there are three types of neural-fuzzy networks (Nauck et al 1997)
existing and are named as Co-operative Neuro-Fuzzy system (Kosko 1992),
Concurrent Neuro-Fuzzy system and Hybrid Neuro-Fuzzy system. A Hybrid
49
Neuro-fuzzy system shall be developed in different ways with each researcher
has defined their own architecture and most particularly are FALCON or
Fuzzy Adaptive Learning Control Network, ANFIS or Adaptive-Network
based Fuzzy Inference System (Abdul Ofoli and Ahamed Rubaai 2004), (Jang
et al 1997), (Principe et al 2000), GARIC or Generalized Approximate
Reasoning based Intelligence Control, NEFCON or Neuronal Fuzzy
Controller; ANFIS is an adaptive network in advance of the different kinds of
feed-forward neural network with supervised learning capability; it is applied
in applications such as non-linear function modelling (Jang 1991,1993), time
series prediction (Jones et al 1990, Jang 1993, Jang and Sun 1993), on-line
parameter identification for control systems (Jang 1993) and design of Fuzzy
controller (Jang 1992). Networks similar to such adaptive network were also
proposed independently by C.T. Lin and [Link] (Lin and George Lee
1991), (Lin and George Lee 1996) and [Link] and [Link] (Wang and
Mendel 1992). However, a major drawback is that there is much freedom in
the choice of structural implementation and it is difficult to decide how
complex a structure is necessary for the desired control to achieve better
system performance (Abdul Ofoli and Ahamed Rubaai 2004).
4.2.1 Architecture of ANFIS
It is assumed that the fuzzy inference system under consideration
has two inputs x and y and an output z for which Type: 3 [Link] and
[Link]’s fuzzy rules are used. For a first-order [Link]’s model (Yager
and Filev 2002), a typical rule base with two fuzzy rules is expressed as:
Rule 1: IF x is A1 and y is B1 THEN f1 = p1x + q1 y + r1
Rule 2: IF x is A2 and y is B2 THEN f2 = p2x + q2 y + r2
Rule 3: IF x is A3 and y is B3 THEN f3 = p3x + q3 y + r3
Rule 4: IF x is A4 and y is B4 THEN f4 = p4x + q4 y + r4
50
The Takagi-Sugeno fuzzy model has the main advantage to model a
system accurately; either globally or locally (Quah and Quek 2006). The
accurate global learning ability motivates the practical applications of this
model in non-linear system estimation (Yen and Langari 1999); the local
learning ability provides a course of interpretability of the local models in the
localized subspaces (Johnson and Babuska 2003, Johnson et al 2000, Yen
et al 1998). Figure 4.3 illustrates a two-input first-order [Link]’s fuzzy
model with two IF-THEN rules and Figure 4.4 clearly depicts the equivalent
architecture of such ANFIS.
Figure 4.3 A two-input first-order [Link]’s Fuzzy model with two rules
51
Figure 4.4 ANFIS architecture
where nodes have the similar function; Layer1 is the input linguistic layer,
Layer2 is conjunction layer, Layer3 is normalization layer, Layer4 is rule or
functional layer and Layer5 is summation layer. It is noted that the structure
of this adaptive network is not unique and hence layers 3 and 4 shall be
combined to obtain an equivalent network with only four layers.
An advantage is that it uses a hybrid learning procedure for estimation of the
premise and consequent parameters (Tahmasebi and Hezarkhani 2010).
