277

Analysis and optimal control of pulse width modulation

feedback systems

C-H Hsieh1 and J-H Chou2*

1Department of Automation Engineering, Chien-kuo Institute of Technology, Changhua, Taiwan
2Department of Mechanical and Automation Engineering, National Kaohsiung First University of Science and
Technology, Kaohsiung, Taiwan

Abstract: This paper considers the analysis and optimization problems of a class of pulse width
modulation (PWM ) feedback systems by using continuous orthogonal functions. Two algorithms
only involving algebraic computation are derived to analyse PWM feedback systems and to
design both the optimal control gains and the optimal PWM patterns of PWM feedback systems
under the criterion of minimizing a quadratic performance index. The illustrated examples show
that the proposed algorithms based on continuous orthogonal functions can obtain satisfactory
results.

Keywords: PWM feedback systems, analysis, optimal control, continuous orthogonal functions,
algebraic algorithms

1 INTRODUCTION references therein). The main characteristic of such tech-

niques is that they convert these problems to those of
Pulse width modulation (PWM ) has been widely used in solving a system of algebraic equations, and hence the
electronic and electrical systems including attitude con- solution and optimization procedures are either greatly
trol systems, adaptive control systems, signal processing, reduced or much simplified accordingly. However, to the
power control systems, and so forth [1]. The reasons authors’ knowledge, among these presented approaches
for the wide applicability of pulse width modulators are of orthogonal functions, only Patra and Rao [4] adopted
that the control variable has only two or three values orthogonal functions to analyse power electronics drive
such that the realization of PWM control is simple, and systems. Patra and Rao [4] have applied general hybrid
the pulse width modulators can process large signals orthogonal functions to analyse silicon controlled rectifier
with high efficiency and low sensitivity to noise. The (SCR) drive systems. However, they did not investigate
PWM feedback systems belong to the field of power the design problems of the power electronics drive systems
electronics having mixed features of continuity and by using orthogonal functions. Because the switching
jumps. The rather unique and inherent non-linear and constant of the PWM pattern depends on the input
discontinuous characteristics make the analysis and design signal of the PWM device and is not a fixed number,
of the power electronics field much more difficult [2–4]. analysing PWM feedback systems is more difficult than
On the other hand, both piecewise and continuous ortho- analysing SCR-controlled drive systems. Therefore, the
gonal functions have received considerable attention in first purpose of this paper is to analyse PWM feedback
recent years as convenient and sharp tools to deal with systems by using continuous orthogonal functions. The
the analysis and optimization problems of dynamic second purpose is to design both the optimal gains and
systems (see, for example, references [4] to [17] and the optimal PWM patterns of PWM feedback systems
by integrating the Taguchi–real-coded genetic algorithm
(Taguchi–RGA) [18, 19] and the continuous orthogonal
The MS was received on 9 June 2003 and was accepted after revision function approach.
for publication on 27 January 2004. This paper is organized as follows. In section 2, the
* Corresponding author: Department of Mechanical and Automation
Engineering, National Kaohsiung First University of Science and analysis of PWM feedback systems by using continuous
Technology, 1 University Road, Yenchao, Kaohsiung 824, Taiwan. orthogonal functions is presented. Two examples are
I06203 © IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

278 C-H HSIEH AND J-H CHOU

also given in this section to demonstrate the availability function representations as

of the proposed algorithm of analysis. Based on the
analysis algorithm presented in section 2, an optimization m−1
x(t)= ∑ x w (t)=xTw(t) (4)
algorithm incorporated with the Taguchi–RGA operation i i
i=0
is derived in section 3 to synthesize both the optimal
m−1
control gains and the optimal PWM patterns under u(t)= ∑ u w (t)=uTw(t) (5)
i i
the criterion of minimizing a quadratic cost functional. i=0
Finally, section 4 offers some conclusions. m−1
A(t)= ∑ A w (t) (6)
i i
i=0
m−1
B(t)= ∑ B w (t) (7)
2 ANALYSIS OF PWM FEEDBACK SYSTEMS i i
i=0
where xT=[ x , x , x , …, x ] is the coefficient matrix,
Consider a PWM feedback system as shown in Fig. 1. 0 1 2 m−1
uT=[u , u , u , …, u ] is the coefficient matrix,
The dynamic equations of the plant are 0 1 2 m−1
wT(t)=[w (t), w (t), …, w (t)] is the orthogonal
0 1 m−1
ẋ(t)=A(t)x(t)+B(t)u(t) basis vector, w (t) (i=0, 1, 2, …, m−1) denote the con-
i
tinuous orthogonal functions, m is the number of terms
y(t)=Cx(t) required for the continuous orthogonal functions, and
x , u , A and B are coefficient matrices of appropriate
(1) i i i i
dimensions. Theoretically, the larger the value of m, the
with the initial state x(0), where x(t)µRn is the state more accurate are the approximate solutions. According
vector, u(t)µR is the control input and y(t)µR is the to previous research experience of the second author
output. Assume that the time interval of interest, [0, T ], [6–10] mµ[4, 8] is generally enough for obtaining
f
is divided into p intervals, i.e. T =pT , where T is the satisfactorily accurate solutions.
f s s Integration of equations (1) from t=t to t yields
sampling period. The operation of the PWM device 0
between its input e(t) and output u(t) is defined as

P P
follows: t t
x(t)−x(t )= A(t)x(t) dt+ B(t)u(t) dt (8)
0
t0 t0
If |e[(q−1)T ]|=c T , then
s q s
where t denotes t or t (q=1, 2, …, p). Now,
0 0q cq
t ∏t<t ;
G
M sgn {e[(q−1)T ]}, using the following integral property of orthogonal
u(t)= s 0q cq (2)
0, t ∏t<t functions [8]
cq fq

P
If |e[(q−1)T ]|T , then t
s s w(t) dt=Hw(t) (9)
t0
u(t)=M sgn {e[(q−1)T ]}, t ∏t<t (3)
s 0q fq
and applying equations (4) to (7), equation (8) can be
where MµR is the amplitude of the PWM device, shown as
q=1, 2, …, p, t =(q−1)T , t =(q−1+c )T , t =
0q s cq q s fq xTw(t)−[x(t ), 0, 0, …, 0]w(t)=aTHw(t)+bTHw(t)
qT , c denotes the switching constant of the PWM 0
s q
pattern for the qth sampling interval and 0∏c <1.
q (10)
Assume that all elements of x(t), u(t), A(t) and B(t) are
absolutely integrable in the interval [t , t ] or [t , t ]. or
0q cq cq fq
Then, for each time interval [t , t ] or [t , t ], they
0q cq cq fq
can be developed into truncated continuous orthogonal xT−[x(t ), 0, 0, …, 0]=aTH+bTH (11)
0

Fig. 1 PWM feedback system

Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering I06203 © IMechE 2004

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

 ANALYSIS AND OPTIMAL CONTROL OF PWM FEEDBACK SYSTEMS 279

where H is the operational matrix of integration [8] and denotes the product constant of two continuous ortho-
can be denoted in general form as gonal functions [9, 10].
Equations (14) and (15) can be rewritten as
t b̃0 ã
1
0 … 0 0 0 u
N N â=Âx̂ and b̂=B̂û (17)
N b̃1 d̃ ã … 0 0 0 N
1 2
N b̃ c̃ d̃ … 0 0 0 N
N 2 1 2 N where â=[aT , aT , …, aT ]T, x̂=[xT , xT , …, xT ]T,
0 1 m−1 0 1 m−1
H= N e e e e e e e N b̂ = [bT , bT , …, bT ]T and û = [uT , uT , …, uT ]T,
N N 0 1 m−1 0 1 m−1
and
N b̃m−3 … … … d̃m−3 ãm−2 0 N
N b̃ … … … c̃ d̃ ã N
N m−2 m−3 m−2 m−1 N t m−1 m−1 m−1 u
N ∑ Ai si00 ∑ As
i i10
… ∑ As
i i,m−1,0 N
v b̃ … … … 0 c̃ d̃ w N i=0 i=0 i=0 N
m−1 m−2 m−1
N m−1 m−1 m−1 N
(12) N ∑ As ∑ As … ∑ As N
Â= N i=0 i i01 i i11 i i,m−1,1 N
i=0 i=0
N N
in which the elements ã , b̃ , c̃ and d̃ (i=1, 2, …, m−1) N e e e e N
i i i i
of H depend on the choice of the orthogonal basis vector N N
m−1
N ∑ As m−1 m−1 N
w(t). Until now, elements of H have been derived for ∑ As … ∑ As
v i i,0,m−1 i i,1,m−1 i i,m−1,m−1 w
various types of continuous orthogonal function [8]. For i=0 i=0 i=0
example, if the Chebychev function is employed, the (18)
constant matrix H is expressed as

t 1 −1 u
0 … 0 0 0
N 2 2 N
N N
N 1 −1 N
0 … 0 0 0
N 8 8 N
N N
N −1 1 N
N 0 … 0 0 0 N
6 4
N N
H=(t −t ) N −1 1 N (13)
f 0 N 0 … 0 0 0 N
16 8
N N
N e e e e e e e N
N N
N −1 1 −1 N
N 2(m−1)(m−3) 0 0 …
4(m−3)
0
4(m−1) N
N N
N −1 1 N
v 2m(m−2) 0 0 … 0 0 w
4(m−2)

within t ∏t∏t . In equations (10) and (11), aT=

0 f
[a , a , a , …, a ] and bT=[b , b , b , …, b ] with t m−1 m−1 m−1 u
0 1 2 m−1 0 1 2 m−1 N ∑ Bi si00 ∑ Bs … ∑ Bs
i i10 i i,m−1,0 N
N i=0 i=0 i=0 N
m−1 m−1 N m−1 N
a = ∑ ∑ Axs , k=0, 1, 2, …, m−1 m−1 m−1
k i j ijk N ∑ Bs ∑ Bs … ∑ Bs N
i=0 j=0 B̂= N i=0 i i01 i=0
i i11
i=0
i i,m−1,1 N
(14) N N
N e e e e N
N N
m−1 m−1 N m−1 m−1 m−1 N
b = ∑ ∑ Bus , k=0, 1, 2, …, m−1 ∑ Bs ∑ Bs … ∑ Bs
k i j ijk v i i,0,m−1 i i,1,m−1 i i,m−1,m−1 w
i=0 j=0 i=0 i=0 i=0
(15) (19)

where Making use of the Kronecker product, equation (11)

can be rewritten as

P
1 tf w (t)w (t)w (t)
s = i j k dt (16)
ijk p √(t −t )(t−t )−(t−t )2 x̂−x̂ =(HTEI )â+(HTEI )b̂ (20)
t0 f 0 0 0 0 n n
I06203 © IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

280 C-H HSIEH AND J-H CHOU

where I is the n×n identity matrix, E denotes the Example 1

n
Kronecker product [20] and
Consider the system in Fig. 1 with the following dynamic
x̂ =[xT(t ), 0T, 0T, …, 0T]T (21) equations:
0 0

C D C DC D C D
ẋ (t) 0 1 x (t) 0
Furthermore, substituting equations (17) into equation 1 = 1 + u(t)
(20) yields ẋ (t) 0 −0.5 x (t) 1
2 2

C D
x̂−x̂ =(HTEI )(Âx̂)+(HTEI )(B̂û) x (t)
0 n n
(22) y(t)=[0 1] 1
x (t)
Finally, the solution of x̂ is achieved as 2
(24)
x̂=(I −(HTEI )Â)−1(x̂ +(HTEI )B̂û) (23) Assume that x (0)=0, x (0)=0, T =1, M=1 and
mn n 0 n 1 2 s
T =6, and that the input h(t) is the unit step function.
where I denotes the mn×mn identity matrix. f
mn The output response of the PWM feedback system with
equations (24) by using the continuous orthogonal
Remark function method is shown in Fig. 2, where the type of
continuous orthogonal function considered in this example
The proposed solution formula of x̂ in equation (23) is is the Chebyshev series. The exact solution and the output
similar to the single-segment multiple-term formula of responses by using the piecewise orthogonal function
general hybrid orthogonal functions given by Patra and approach (i.e. the improved block pulse function approach
Rao [4]. and the Haar wavelet approach) are also shown in Fig. 2.
The switching constant c of the PWM pattern depends
q From Fig. 2 it can be seen that the continuous ortho-
on the input signal e(t) of the PWM device and is not gonal function method may give better results than the
a fixed number, so, based on the above derivation, the piecewise orthogonal function approach.
following algorithm only involving algebraic computation
is proposed for calculating the solution of the PWM Example 2
feedback system shown in Fig. 1, the dynamic equations
of the plant being given in equations (1). The dynamic equations of the plant in Fig. 1 are

C D C DC D C D
ẋ (t) 1 e−t x (t) 0
Step 1. Give x(0), sampling period T , amplitude M and 1 = 1 + u(t)
s ẋ (t) −e−t 1 x (t)
input h(t). 2 2
1
Step 2. Compute y(0)=Cx(0) and e(0)=h(0)−y(0); set

C D
x (t)
q=1. y(t)=[1 0] 1
Step 3. Determine c and û for the time interval [t , t ] x (t)
q 0q fq 2
according to equations (2) and (3).
(25)
Step 4. If c <1, then
q
(a) set t =t , calculate  and B̂ from equations (18) Assume that x (0)=0, x (0)=0, T =0.25, M=1,
0 0q 1 2 s
and (19) and calculate x̂ from equation (23); T =1.5 and h(t)=0.35.
f
(b) set t =t and compute x(t )=ŵT (t )x̂T, where Based on the Chebyshev series, the output response
f cq f n f
ŵT (t )=[I w (t ), I w (t ), …, I w (t )]; of the PWM feedback system with equations (25) is
n f n 0 f n 1 f n n−1 f
(c) set t =t and let x(t )=x(t ). Calculate  and B̂ shown in Fig. 3. Example 1 has shown that the con-
0 cq 0 f
from equations (18) and (19) again, and calculate tinuous orthogonal function method can obtain better
x̂ from equation (23); results than the piecewise orthogonal function approach.
(d) set t =t and compute x(t )=ŵT (t )x̂T. Let t =t Therefore, in this example, it is only necessary to com-
f fq f n f 0 f
and x(t )=x(t ), and go to step 5. pare the result with the exact solution. It is obvious from
0 f
Else if c 1, then Fig. 3 that the presented continuous orthogonal function
q
(a) set t =t , calculate  and B̂ from equations (18) method is suitable for analysing the PWM feedback
0 0q
and (19) and calculate x̂ from equation (23); system shown in Fig. 1.
(b) set t =t and compute x(t )=ŵT (t )x̂T. Let t =t
f fq f n f 0 f
and x(t )=x(t ), and go to step 5.
0 f
Step 5. Compute y(t )=Cx(t ) and e(t )=h(t )−y(t ). 3 OPTIMAL CONTROL OF PWM FEEDBACK
f f f f f SYSTEMS
Step 6. Set q=q+1. If q>p, then stop; otherwise go
to step 3.
Consider the PWM feedback system shown in Fig. 4. The
In what follows, two examples are given to compare the problem is to find both the optimal control gain, K , and
q
results obtained by the method presented in this paper the optimal switching constant, c , at the qth sampling
q
with their exact solutions. interval, for q=1, 2, …, p, such that the following
Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering I06203 © IMechE 2004

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

 ANALYSIS AND OPTIMAL CONTROL OF PWM FEEDBACK SYSTEMS 281

Fig. 2 Output responses of example 1 by using piecewise orthogonal functions (improved block pulse
functions and Haar wavelets) and continuous orthogonal functions (Chebyshev series), with the
exact solution

Fig. 3 Output response of example 2 by using continuous orthogonal functions (Chebyshev series) with the
exact solution

Fig. 4 Optimal control of the PWM feedback system

I06203 © IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

282 C-H HSIEH AND J-H CHOU

performance index is minimized: Step 7. Set t =t and compute x(t )=x̃T w(t ).
f fq f q f
Step 8. Let t =t and x(t )=x(t ).
0 f 0 f

CP
p tcq Step 9. Set q=q+1. If q>p, then stop; otherwise go
J= ∑ [Qe2(t)+Ru2(t)] dt to step 3.
q=1 t0q
From the above algorithm, it is obvious that, if c
P D
tfq q
+ [Qe2(t)+Ru2(t)] dt (26) (q=1, 2, …, p) and M are specified, then x: , x̃ , u:
q q q
tcq and ũ (q=1, 2, …, p) can be determined and the
q
value of the performance index in equation (27)
where e(t)=h(t)−y(t), Q is a non-negative real number
corresponding to this set of c (q=1, 2, …, p) and M
and R is a positive real number. Substituting the truncated q
is obtained. Therefore, the value of the performance
continuous orthogonal function representations of x(t)
index in equation (27) is actually dependent on c
and u(t) [i.e. equations (4) and (5)] into equation (26), q
(q=1, 2, …, p) and M, i.e.
the performance index J in equation (26) becomes
J=F(c , c , …, c , M) (29)
p 1 2 p
J= ∑ [c T (trace ( W(e: Qe: T +u: Ru: T ))) Hence, the optimal control problem of the PWM feedback
q s q q q q
q=1 system is to search for the optimal c (q=1, 2, …, p)
+(1−c )T (trace ( W(ẽ QẽT +ũ RũT )))] q
q s q q q q and M such that the performance index in equation (26)
(27) is minimized. This is equivalent to the problem
min J=F(c , c , …, c , M) (30a)
where e: =h: −x: CT, ẽ =h̃ −x̃ CT; h: , x: and u: 1 2 p
q q q q q q q q q subject to
denote respectively the continuous orthogonal function
coefficient matrices of h(t), x(t) [i.e. equation (4)]
0∏c <1 (q=1, 2, …, p) and 0∏M∏c
and u(t) [i.e. equation (5)] for the time interval [t , t ]; q
0q cq
h̃ , x̃ and ũ denote respectively the continuous (30b)
q q q
orthogonal function coefficient matrices of h(t), x(t)
where c is a positive real number denoting the limitation
[i.e. equation (4)] and u(t) [i.e. equation (5)] for the time
of M given from the practical consideration.
interval [t , t ]; and W denotes the product integration
cq fq Equation (30a) is a non-linear function. For the
matrix of two continuous orthogonal function basis
optimization problem of a non-linear function with
vectors [6, 10] and is defined as
continuous parameters, real-coded genetic algorithms
( RGAs), i.e. using floating-point representation as
P
tf
W= w(t)wT(t) dt (28) genes, have been shown to have noticeable performance,
t0 whereas the bit-string representation adopted in classical
genetic algorithms will result in a loss of precision
Now, if one set of {c , c , …, c } with 0∏c <1 [21–23]. On the other hand, it is well known that the
1 2 p q
(q=1, 2, …, p) and the amplitude M of the PWM configuration of operating parameters (for example,
device are given, then x: and x̃ for q=1, 2, …, p population size, crossover rate, mutation rate and cross-
q q
can be calculated from the following algorithm only over operator) in RGA has long been a troublesome
involving algebraic computation. problem [23]. The Taguchi method [24] is a kind of
experimental design method broadly used in quality
Step 1. Give x(0), sampling period T and input h(t).
s engineering to obtain better quality and/or larger pro-
Set q=1. ductivity of industry products. Therefore, in this paper,
Step 2. Compute y(t )=Cx(t ) and e(t )=h(t )−
0q 0q 0q 0q the Taguchi–RGA method [18, 19], which integrates
y(t ).
0q both the Taguchi method and the RGA, is employed
Step 3. Choose K >0 so that K |e[t ]|=c T
q q 0q q s to search for the optimal solution of the problem in
equations (30). Here, the purpose of using the Taguchi
u: =[M sign {e[t ]}, 0, 0, …, 0]T for t ∏t<t
q 0q 0q cq method is to find the optimal combination of the
ũ =[0, 0, 0, …, 0]T for t ∏t<t operating parameters in the RGA such that the efficiency
q cq fq of the RGA can be improved.
Step 4. Set t =t , calculate  and B̂ from equations A brief description of the Taguchi–RGA method
0 0q [18, 19] is as follows.
(18) and (19) and calculate x: from equation (23)
q
[i.e. x̂ in equation (23)]. Step 1. Define the signal-to-noise ratio (SNR), g, for
Step 5. Set t =t and compute x(t )=x: T w(t ).
f cq f q f the Taguchi matrix experiment. In this paper, g is
Step 6. Set t =t and let x(t )=x(t ). Calculate  and B̂
0 cq 0 f expressed in decibels by
from equations (18) and (19) again, and calculate x̃
q
from equation (23) [i.e. x̂ in equation (23)]. g=−10 log ( f ) (31)
Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering I06203 © IMechE 2004

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

 ANALYSIS AND OPTIMAL CONTROL OF PWM FEEDBACK SYSTEMS 283

in which f is the fitness function defined for the RGA. crossed. Then, an offspring V∞={v∞ , …, v∞ , …, v∞ } is
1 i n
The fitness function adopted in this paper is described produced by the crossover of V 1 and V 2, where v∞ is
i
as J, where J is the performance index in equation (27). randomly selected in the interval [v −Dv a, v: +Dv a],
i i i i
Step 2. Each of the four operating parameters (factors) v =min (v1 , v2 ), v: =max (v1 , v2 ), Dv =v: −v and
i i i i i i i i i
of the RGA, i.e. population size, crossover rate, aµ(0, 1) is the parameter of the BLX_a crossover
mutation rate and the parameter a of the BLX_a operator.
crossover operator, is assigned three distinct levels, Step 5. With respect to the mutation operation, the
and the levels of each operating parameter are listed non-uniform mutation [22] is adopted in this paper.
in Table 1. The purpose of the Taguchi matrix experi- For a given parent V={v , v , …, v }, its element v
1 2 n k
ment is to determine the best level for each operating is mutated if a randomly generated number is smaller
parameter such that g is maximized. The matrix than a predefined mutation rate P , and then
m
experiment, which is a standard L array, selected for the mutated element v∞ in the generated offspring
9 i
this paper is given in Table 2 to study the operating V∞={v∞ , …, v∞ , …, v∞ } is randomly selected from two
1 i n
parameter choice of the RGA with four three-level possible choices:
factors.
Step 3. Each row in Table 2 indicates a set of operating v∞ =v +D(s, v: −v ) or v∞ =v −D(s, v −v )
i i i i i i i i
parameters of the RGA. For each RGA operation, a (33)
set of parameters in the ith (i=1–9) row is employed,
and the initial value of the row index i is 1. That is, where v: and v are the upper bound and lower bound
i i
the first set of operating parameters of the RGA is of v . The value of function D(s, d) is in the range
i
A , B , C and D . Like the general genetic algorithm, [0, d] such that the value of D(s, d) approaches 0 as
1 1 1 1
firstly, the initial population with chromosomes of the generation number, s, increases, and D(s, d)=
the form dj(1−s/G)h, where j is a random number from
[0, 1], G is the maximal generation number and h is a
U={c , c , …, c , M} (32)
1 2 p parameter determining the degree of non-uniformity.
[where 0∏c ∏1 (q=1, 2,…, p) and 0∏M∏C ] is Step 6. In this step, the fitness value of each chromosome
q in the parameter set of equation (32) is evaluated
randomly generated.
Step 4. For the crossover operation, the BLX_a through one practical simulation by using the control
crossover operator [18, 19, 21] is applied to the parameters in the chromosome.
parameters c and M in U. Let V 1={v1 ,·v1 , …, v1 } Step 7. The selection philosophy adopted in the repro-
q 1 2 n duction process of this step is the (m+l) selection
and V 2={v2 , v2 , …, v2 } be two chromosomes to be
1 2 n method [23] in which m parents and l offspring com-
pete for survival and the m best out of both the old
parents and the old offspring are selected as parents
Table 1 Levels of each operating parameter of the next generation. If the termination condition
(maximum number of iteration) of the RGA is not
Level
reached, then go to step 8; otherwise go to step 4.
Factor 1 2 3 Step 8. The lowest fitness value in the RGA operation
from step 4 to step 7 is utilized to calculate the
A (population size) 20 (A ) 40 (A ) 60 (A )
1 2 3 corresponding g in Table 2. Add the row index i by 1,
B (crossover rate) 0.5 (B )
1
0.7 (B )
2
0. 9 (B )
3 i
C (mutation rate) 0.01 (C )
1
0.02 (C )
2
0.03 (C )
3
i.e. i=i+1. If the row index i is equal to 9, then go
D (a) 0.25 (D ) 0.5 (D ) 0.75 (D ) to step 9; otherwise go to step 3.
1 2 3
Step 9. Throughout the whole Taguchi matrix experiment,
i.e. nine RGA operations, g (i=1–9) are acquired.
i
The overall mean of the SNR for the experiments
Table 2 Matrix experiment using an L orthogonal array
9 defined in Table 2 is given by g: = 1 W 9 g . Table 3
9 k=1 k
Factor
Experiment
number A B C D g (SNR)
Table 3 SNR of each operating parameter and level
1 1 1 1 1 −18.4607 (g )
1
2 1 2 2 2 −18.2373 (g ) Level
2
3 1 3 3 3 −19.3741 (g )
3
4 2 1 2 3 −18.3605 (g ) Factor 1 2 3
4
5 2 2 3 1 −18.0029 (g )
5
6 2 3 1 2 −18.2453 (g ) A −18.6907 (g ) −18.2029 (g ) −18.3199 (g )
6 A A A
7 3 1 3 2 −17.7656 (g ) B −18.1956 (g 1) −18.1726 (g 2) −18.8453 (g 3)
7 B B B
8 3 2 1 3 −18.2775 (g ) C −18.3279 (g 1 ) −18.5048 (g 2 ) −18.3809 (g 3 )
8 C C C
9 3 3 2 1 −18.9165 (g ) D −18.4601 (g 1 ) −18.0827 (g 2 ) −18.6707 (g 3 )
9 D1 D2 D3

I06203 © IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

284 C-H HSIEH AND J-H CHOU

lists the main effects of each operating parameter, in Let c* (q=1, 2, …, p) and M* be the optimal solutions
q
which obtained from the above Taguchi–RGA operation. Then
the optimal control gains K* (q=1, 2, …, p) can be
g = 1 (g +g +g ), g = 1 (g +g +g ) q
A1 3 1 2 3 A2 3 4 5 6 obtained in step 3 of the algorithm involving only the
1 1 algebraic computation presented in the preceding part
g = (g +g +g ), g = (g +g +g )
A3 3 7 8 9 B1 3 1 4 7 of this section.
g = 1 (g +g +g ), g = 1 (g +g +g )
B2 3 2 5 8 B3 3 3 6 9
Example 3
g = 1 (g +g +g ), g = 1 (g +g +g )
C1 3 1 6 8 C2 3 2 4 9 Consider the PWM feedback system in Fig. 4 with the
g = 1 (g +g +g ), g = 1 (g +g +g ) following dynamic equations:
C3 3 3 5 7 D1 3 1 5 9
ẋ (t) x (t)
C D C DC D C D
1
g = (g +g +g ), 1
g = (g +g +g ) 1
0 1
1
0
D2 3 2 6 7 D3 3 3 4 8 = + u(t)
ẋ (t) −4 −1 x (t) 1
(34) 2 2
x (t)
From Table 3 it is possible to predict the optimal
level for each factor having the highest value of g.
y(t)=[0 1] 1
x (t)
2
C D
For example, if the highest values of g corresponding (35)
to factors A, B, C and D are g , g , g and g
A B2 C2 D Assume that x (0)=3, x (0)=−2, T =0.5, M=1,
respectively, then the best levels of2 A, B, C and D are3 1 2 s
A , B , C and D respectively. Based on the Taguchi T =10 and h(t)=0. The performance index is
2 2 2 3 f
matrix experiment, it can be concluded that the
CP
20 tcq
setting A B C D is the optimal combination of the J= ∑ [Qe2(t)+Ru2(t)] dt
2 2 2 3
operating parameters. q=1 t0q

P D
Step 10. As described in step 9, if the optimal com- t
bination is A B C D , it is possible to predict the + fq [Qe2(t)+Ru2(t)] dt
2 2 2 3
value of the SNR under the optimal setting, denoted tcq
by g , and g =g +g +g +g −3g: . Now, where Q=1, R=1, t =0.5(q−1), t =0.5(q−1+c )
opt opt A2 B2 C D3 0q cq q
using the predicted optimal setting2 obtained in step 9 and t =0.5q. Based on the Chebyshev series with
fq
to perform the RGA operation (steps 4 to 7) once m=4, after application of the presented search algorithm
again, the chromosome with the lowest fitness value with the Taguchi–RGA operation [18, 19], the optimal
is selected and decoded to yield the optimal solutions. switching constant c* (q=1, 2, …, 20) of the PWM
q

Fig. 5 History of minimum fitness value against generation

Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering I06203 © IMechE 2004

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

 ANALYSIS AND OPTIMAL CONTROL OF PWM FEEDBACK SYSTEMS 285

device can be obtained, and thus the optimal control After using the Taguchi–RGA operation with the
gains K* (q=1, 2, …, 20). The optimal c* and the operating parameters shown in Table 1, the calculated
q q
optimal K* (q=1, 2, …, 20) are listed in the following results are shown in Tables 2 and 3. Then, it can be con-
q
table: cluded that the predicted optimal setting is A B C D .
2 2 1 2
Figure 5 shows the graph of the minimum fitness value
q 1 2 3 4 5 6 7 8 9 10
against generation, which uses the predicted optimal
K* 0.250 0.115 0.122 1.210 0.150 0.199 0.031 0.001 0.433 0.006 setting A B C D (i.e. the population size is 40, the
c*
q
1 0.968 0.609 0.405 0.414 0.460 0.013 0.001 0.385 0.001
2 2 1 2
q crossover rate is 0.7, the mutation rate is 0.01 and the
q 11 12 13 14 15 16 17 18 19 20 parameter a of the BLX_a crossover operator is 0.5).
K* 0.269 0.729 0.013 0.465 4.289 44.515 6.416 0.008 1.502 0.019
q The output response of the optimal PWM feedback
c* 0.035 0.155 0.001 0.017 0.081 0.170 0.105 0.001 0.080 0.001
q system with equations (35) is shown in Fig. 6. The

Fig. 6 Output response of example 3 by using continuous orthogonal functions (Chebyshev series)

Fig. 7 Optimal PWM patterns of u(t) for example 3

I06203 © IMechE 2004 Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016

286 C-H HSIEH AND J-H CHOU

optimal PWM patterns are also obtained and shown 5 Rao, G. P. Piecewise Constant Orthogonal Functions and
in Fig. 7. Their Applications to Systems and Control, 1983 (Springer-
Verlag, Berlin, Germany).
6 Horng, I. R. and Chou, J. H. The design of optimal
observers via shifted Chebyshev polynomials. Int. J.
4 CONCLUSIONS Control, 1985, 41, 549–556.
7 Horng, I. R. and Chou, J. H. Analysis, parameter
Based on continuous orthogonal functions, two algorithms estimation and optimal control of time-delay systems via
have been presented in this paper. One algorithm is Chebyshev series. Int. J. Control, 1985, 41, 1221–1234.
used to analyse PWM feedback systems. Example 1 has 8 Chou, J. H. and Horng, I. R. State estimation using
shown that the analysis algorithm based on continuous continuous orthogonal functions. Int. J. Syst. Sci., 1986,
orthogonal functions can obtain better results than 17, 1261–1267.
the analysis algorithm based on piecewise orthogonal 9 Chou, J. H. and Horng, I. R. Shifted-Chebyshev-series
functions. The reason for this is that, since the signals analysis and identification of time-varying bilinear systems.
Int. J. Control, 1986, 43, 129–137.
over each segment [t , t ] or [t , t ] are continuous,
0q cq cq fq 10 Chou, J. H. Application of Legendre series to the optimal
the continuous orthogonal functions naturally do better
control of integrodifferential equations. Int. J. Control,
than the piecewise orthogonal functions. The other 1987, 45, 269–277.
algorithm incorporated with the Taguchi–RGA operation 11 Datta, K. B. and Mohan, B. M. Orthogonal Functions in
is used to synthesize both the optimal control gains Systems and Control, 1995 ( World Scientific, Singapore).
and the optimal PWM patterns of PWM feedback sys- 12 Elnagar, G. N. and Razzaghi, M. A collocation-type
tems under the criterion of minimizing a quadratic cost method for linear quadratic optimal control problems.
functional. Both presented algorithms of analysis and Optimal Control Applic. and Meth., 1997, 18, 227–235.
design involve only algebraic computation, so the solution 13 Chen, C. F. and Hsiao, C. H. Wavelet approach to
and optimization procedures of PWM feedback systems optimizing dynamic systems. IEE Proc. Control Theory
can be either greatly reduced or much simplified accord- Applic., 1999, 146, 213–219.
ingly. On the other hand, in this paper, equation (1) 14 Hsiao, C. H. and Wang, W. J. Optimal control of linear
assumes that the plant is linear. However, if a quasi- time-varying systems via Haar wavelets. J. Optimization
Theory and Applic., 1999, 103, 641–655.
linearization technique [17] is used, the proposed
15 Hsiao, C. H. and Wang, W. J. Haar wavelet approach to
algorithms could be extended to PWM feedback systems
nonlinear stiff systems. Math. and Computers in Simulation,
with non-linear plants. This extension is now under 2001, 57, 347–353.
investigation. 16 Pacheco, R. P. and Steffen Jr, V. Using orthogonal
functions for identification and sensitivity analysis of
mechanical systems. J. Vibr. Control, 2002, 8, 993–1021.
ACKNOWLEDGEMENT 17 Jaddu, H. Direct solution of nonlinear optimal control
problems using quasilinearization and Chebyshev poly-
This work is in part supported by the National Science nomials. J. Franklin Inst., 2002, 339, 479–498.
Council, Taiwan, Republic of China, under grant number 18 Hsieh, C. H., Chou, J. H. and Wu, Y. J. Optimal grey-
fuzzy gain-scheduler design using Taguchi–HGA method.
NSC91-2213-E270-001.
J. Intell. and Robotic Syst., 2001, 32, 321–345.
19 Hsieh, C. H., Chou, J. H. and Wu, Y. J. Optimal predicted
fuzzy PI gain scheduling controller of constant turning
REFERENCES force systems with fixed metal removal rate. Int. J. Advd
Mfg Technol., 2002, 19, 714–721.
1 Hou, L. and Michel, A. N. Stability analysis of pulse- 20 Barnett, S. Matrix Methods for Engineers and Scientists,
width-modulated feedback systems. Automatica, 2001, 37, 1979 (McGraw-Hill, New York).
1335–1349. 21 Eshelman, L. J. and Schaffer, J. D. Real-coded genetic
2 Wu, S. G. Analysis and design of sampled-data con- algorithms and interval schemata. In Foundation of Genetic
trol systems. PhD dissertation, Department of Power Algorithms—2 (Ed. L. D. Whitley), 1993 (Morgan
Mechanical Engineering, National Tsing-Hua University, Kaufmann, Vermont).
Taiwan, 1987. 22 Michalewicz, Z. Genetic Algorithm+Data Structures=
3 Williamson, D. Digital Control and Implementation, 1991 Evolution Programs, 1994 (Springer, New York).
(Prentice-Hall, Englewood Cliffs, New Jersey). 23 Gen, M. and Cheng, R. W. Genetic Algorithms and
4 Patra, A. and Rao, G. P. General Hybrid Orthogonal Engineering Design, 1997 (John Wiley, New York).
Functions and Their Applications in Systems and Control, 24 Belavendram, N. Quality by Design, 1995 (Prentice-Hall,
1996 (Springer-Verlag, London). Englewood Cliffs, New Jersey).

Proc. Instn Mech. Engrs Vol. 218 Part I: J. Systems and Control Engineering I06203 © IMechE 2004

Downloaded from pii.sagepub.com at NANYANG TECH UNIV LIBRARY on June 5, 2016