ACSE Volume12 Issue2 P1111434342

The International Congress for global Science
and Technology
(ICGST)
ICGST International Journal on Automatic

Control & System Engineering
(ACSE)
Volume (12), Issue (II)

October 2012
wwww.icgst.com
www.icgst-amc.com
www.icgst-ees.com
© ICGST LLC, Delaware, USA, 2012

ACSE Journal
ISSN: Print 1687-4811
ISSN Online 1687-482X
ISSN CD-ROM 1687-4838
© ICGST LLC, Delaware, USA 2012
Table of Contents
Papers Pages
P1111209135,
Tarun Varshney and Satya Sheel,
"PID Control of MIMO Turbo – Generator Plant Based on Genetic 1--6
Algorithm"
P1111217176,
Mohamed F. Hassan and Hala A. Mourad,
"Closed Loop Controller for Stochastic Systems with Uncertain 7--14
Parameters"
P1111210136,
O. Chandra Sekhar, Dr.K.Chandra sekhar Member, IEEE,
"A Novel Nine-Level MPC Inverter for Direct Torque Control Induction 15--22
Motor Drive "
P1111219191,
Akram Abu-aisheh and Sameer Khader,
23--30
"Hybrid MPPT-Controlled LED Illumination Systems"
P1111212142,
G. Sambasiva Rao and K. Chandra Sekhar,
"A sophisticated Space Vector Pulse Width Modulation Signal Generation 31--34
for Nine-Level Inverter system for Dual-Fed Induction Motor Drive"
P1111223205,
H. Bellahsene and I. F. E. Fatani,
"Amélioration du temps de convergence de l’algorithme de calcul du filtre 47--51
optimal par HOS en DMT"
P1111217174,
C. Harikrihsna and T. Bramhananda Reddy and J. Amarnath and S.
Kamakshaiah, 39--46
"Simple and Novel Generalized Scalar PWM Algorithm for Multilevel
Inverter Fed Direct Torque Controlled Induction Motor"
ICGST International Journal on Automatic Control &
System Engineering - (ACSE)
A Publication of the International Congress for

global Science and Technology - (ICGST)
ICGST Editor in Chief: Dr. rer. nat. Ashraf Aboshosha
www.icgst.com, www.icgst-amc.com,
www.icgst-ees.com
editor@icgst.com
ICGST-ACSE Journal, ISSN 1687-4811, Volume 12, Issue 2, Delaware, USA, October 2012
Improved PID MIMO Control of Turbo Generator relying on Genetic

Algorithm
Tarun Varshney, Satya Sheel

Electrical Engineering Department, Motilal National Institute of Technology, Allahabad, India,
t_varshney@yahoo.com, satyasheel@mnnit.ac.in
Abstract the nonlinear SISO Hammerstein model. Maja and Kabra

PID controllers are widely used in industries due to [8] presented a tuning method for multivariable controller
simple functioning and robustness, yet the determination that was based on decoupling control and GA. As this
of controller gains is a big challenge more so when the method tends to achieve the shortest direct paths through
plant to be control is multi input multi output (MIMO) the system (without inverting the model), it usually
and nonlinear in nature. The aim of this paper is to string results in low energy consumption which can further be
together the qualities of conventional PID controller and directly controlled with pole placement. Jih-Gau Juang et
a recent technique Genetic Algorithm (GA). In this work al. [12] used GA to tune the PID parameters to control a
controller parameters are adjusted by means of fitness laboratory helicopter called twin rotor system. All
function that can evaluate the optimum PID parameters parameters of the controller are obtained by a real-value-
based on system error. Simulation study has been carried type genetic algorithm (RGA) with an integral of time
out for a turbo- generator and the results have been multiplied by the square error criterion. Jin-Sung Kim et.
compared with decoupled control and diagonal recurrent al [13] proposed tuning method which was based on
neural network (DRNN) based PID tuning methods. A improved GA. For better performance they defined new
simulation result shows that the proposed control scheme objective function in the sense of root mean square error.
results in smaller controller effort. In addition they made improvement by replacing the
uniform distribution random number generation in
Keywords: Genetic algorithm, PID Tuning, MIMO conventional GA to newly designed hybrid random
system generator composed of Cauchy distribution and linear
congruential generator which provides independent and
1. Introduction different random numbers at each individual steps in
PID controllers are widely used in industries due to their Genetic operations. In many research works GA has been
simple functioning and robustness but determination of widely applied to tune the PID parameters such as a
controller gains has always been a big challenge. Most nonlinear model of thermal process [9], gasifier problem
industrial processes are multivariable and their nonlinear [11], missile control system, DC motor position control,
and interactive behavior imposes difficulties in LTI system, liquid level system [14-17]. Monica
determination of the parameters of controller. In such a Patrascu et al. [18] applied GA to tune the parameters of
situation progressive technique like GA has been shown PID controller for nonlinear model MIMO system. They
to be capable of high performance without experiencing have focused on the initialization procedure of the
the difficulties [1], [10]. GA offers an adaptive heuristic algorithm that can influence the outcome of the
search algorithm based on the evolutionary ideas of optimization procedure and showed that roulette wheel
selection and genetics [2]. It is inspired by Darwin’s selection is better choice for selection. Daniel Carmona
theory about evolution- “Survival of fittest”. GA Morales et.al [20] developed a Graphical user interface
represents an intelligent exploitation of random search system for optimum tuning of multivariable PID
used to solve optimization problems. Takhiro and Sigeru controller. From the educational point of view, this tool
[3] applied GA to tune the PID gains for a single input provides students with the necessary means to
single output (SISO) water bath system and proved that consolidate their knowledge on these control structures.
real number encoding is simple to operate and better In this work GA is chosen to tune the parameters of PID
than bit string encoding method. Mitsukura, Yamamoto controller for a MIMO Turbo- Generator plant which is
and Kaneda [4, 5] proposed a new tuning method of PID highly nonlinear and has strong coupling among
parameters based on the relationship between the PID variables. The simulation results have obtained by the
control law and the generalized minimum variance proposed approach been compared with those obtained
control show the effectiveness of proposed method for by fixed decoupled PID configuration and also with
DRNN based tuning of controller parameters [19].
1
The rest of paper is organized as follows. In the next reproduction operator. Selection means extract a subset
section basic steps of GA are described. In Section 3 of genes from an existing population. Every gene has a
Genetic operator such as Selection, Crossover and meaning so one may extract a kind of quality
mutation are illustrated. Section 4 describes the measurement called as fitness value and by following this
evaluation procedure based on fitness function and quality selection can be done. Fitness function quantifies
fitness value. In section 5 the basic ideas of the tuning the optimality of a solution so that a particular solution
method are presented. Section 6 describes the Turbo- may be ranked against all the other solutions. The
Generator plant considered and simulation studies are function shows the degree of the closeness of the given
presented in Section 7. Concluding remarks are in last solution to the desired solution. Roulette wheel selection,
Section. Rank selection, Boltzmann selection, Steady state
selection and tournament selection are examples of
2. Genetic Algorithm selection operators [18].
GA is a stochastic search method that emulates the B. Crossover
process of natural evolution. It starts with the minimal After the selection, crossover operator takes place and it
knowledge of the correct solution and depends entirely combines two chromosomes (parents) to produce anew
on responses from its environment and genetic operators chromosome (offspring).The centre idea of crossover is
e.g. selection or reproduction, crossover and mutation to that the new generated chromosome may be better than
get the optimal solution [1]. By starting at several
independent points and searching in parallel, the GA
avoids local minima and converges to sub optimal
solutions. The GA is typically initialized with a random
population consisting of between 20-100 individuals.
This population is usually represented by a real valued
number or a binary string called a chromosome.
Individual performance is measured by the fitness
function. The fitness function assigns each individual a
corresponding number called its fitness value. The fitness
value of each chromosome is assessed and a survival of
the fittest strategy is applied. There are three main
genetic operators of the GA, these are known as Selection
or reproduction, crossover and mutation.
The steps involved in implementing the GA are as
follows:
1. Generate an initial, random population of
individuals for a fixed size.
2. Applying it to the process
3. Evaluate the process using fitness function chosen
i.e. IAE plus weighted control efforts (squared)
4. Select the fittest members of the population.
5. Reproduce using a probabilistic method.
6. Implement crossover operation on the reproduced
chromosomes
7. Execute mutation operation with low probability.
8. Repeat step 2 until a predefined convergence
criterion is met. The convergence criterion of a
genetic algorithm is a user-specified condition e.g.
the maximum number of generations or when the
string fitness value exceeds a certain threshold.
Figure 1 Flowchart of Genetic Algorithm
An illustrative flowchart of the GA implementation is
shown in Figure 1
both the parents if it takes the better characteristics from
3. Genetic Operator each of them. It depends on user specified crossover
To maintain the genetic diversity GA uses the operators. probability. It indicates how often crossover is performed.
Genetic diversity is a necessarily procedure for process of A probability of 0% means that the ‘offspring’ will be
evolution. Commonly three main Genetic operators of exact replicas of their ‘parents’ and a probability of 100%
selection or reproduction, crossover and mutation are means that each generation will be composed of entirely
used in implementation of GA. new offspring. One point crossover, two points, uniform,
arithmetic and heuristic are few types of crossover.
A. Selection or Reproduction C. Mutation
Selection is usually the first operator applied on Mutation is the last operation in GA. It is used to
population. The Selection operator is also known as maintain genetic diversity from one generation of a
2
population of chromosome to the next generation. It u(k) = k p (k)x1 + ki (k)x2 + kd (k)x3 (4)
depends upon the user defined mutation probability.
Mutation probability values of around 0.1% or 0.01% are
common, these values represent the probability that a n
certain string will be selected for mutation i.e. for a where x1 ( k ) = e(k) , x2 =  e(k)T ,
k=1
probability of 0.1%; one string in one thousand will be
selected for mutation. Mutation alters one or more gene e(k) - e(k - 1)
x3 = , and kp, ki, kd are proportional, integral,
values in a chromosome from its initial state. This can T
result a fairly new gene value being added to the and derivative gains respectively and T is sampling time.
population and with new gene value the GA is able to get The convergence criterion of a GA is a user-specified
better solution in contrast to previous solution. Mutation condition and termination of tuning of PID parameters
helps to prevent the population from stagnating at any can be done by reaching the maximum number of
local minima. The mutation operators are of many types generations or when the fitness value exceeds a certain
such as flit bit, Boundary, Non uniform, Uniform and threshold. To get the satisfactory dynamic characteristics
Gaussian. IAE has been considered as the fitness function in the
present work. To limit the controlled input from being
4. Process Evaluation Based on Fitness too large, it also accumulates the quadratic term of
controlled input to the fitness function. The expression
Function for modified fitness function can be represented as
A. Fitness function / Performance index
follows
The most crucial action to choose the fitness function to
J =   w1 e(k) + w2 u 2 (k)
n
evaluate the fitness of each chromosome. Performance (5)
indices such as mean square error (MSE), integral time k=1
multiplied by absolute error (ITAE), integral absolute
error (IAE) and integral square error are commonly used
as the fitness function in applying GA [7].
B. Fitness Value
The PID controller is designed to minimize the fitness
function which is function of system error. The smaller
value of fitness function indicates the higher fitness value
of the corresponding chromosome and vice versa, and it
is represented by
1
fitness value = (1)
fitness function
5. Problem Formulation Figure 2 Schematic diagram of PID tuning based on GA

Figure 2 depicts the schematic diagram of the proposed
genetic tuning of PID parameters. The scheme composes 6. The Turbo - Generator Plant
of usual feedback control system which has the full The schematic diagram of laboratory scale turbo-
model of coupled nonlinear multivariable plant and the generator plant is shown in Figure 3, It was developed as
PID controller. The parameters of controller have been an emergency power supply for biological experiments
recursively tuned online by GA. To begin with, we [6]. A compressed centre air supply runs a turbine which
assume that system to be controlled has multiple inputs drives a Y - connected three phase synchronous generator.
u (k )  m X 1 and multiple outputs y (k )  p X 1 and The field of the generator is separately excited by an
external source. A balanced three phase resistance serves
the control objective is to follow a prescribe set points
as a load which can be switched between 1.5 Ω and 2.5 Ω.
r (k )  p X 1 . Let the tracking error be defined by A tacho generator is used for speed measurements. The
percentage of valve opening V (%) and field current iE
e(k) = r(k) - y(k) (2) (mA) are input signals u1(t) and u2(t) respectively. The
turbine speed n (min-1) and the rectified mean value of
and the controlled input of the system is the generated voltage UA (V) are considered as the output
signals y1(t) and y2(t) respectively. The numbers of
machines poles are constant so the frequency f (Hz) of
u (k )  K (k ) X (k ) (3)
the generated unrectified voltage is proportional to speed
n.
where K ( k ) 
mX 3
is the set of controller’s
parameters [kp(k) ki(k) kd(k)] and of and
X (k ) 3 X p is the set of errors which is represented
by [x1(k) x2(k) x3(k)]T.
3
different techniques is also shown. It can be seen that

u1 (t ) ~ V (t ) u1 (t ) ~ iE (t )
with the help of proposed scheme system response
becomes quite faster, rise time becomes shorter and
servo Current smaller controlled input is required to meet the desire
system source response (Figure 5). The optimization process of the
fitness function is shown in Figure 6 which shows the
optimize value of fitness function at each iteration.
Finally optimize PID parameters
balanced  0.1985 0.1458 0.0119 
are K  
GT
load .
0.1166 0.0454 0.0207 
y1 (t ) ~ n(t )
Synchoronous
Turbine
generator All the computation has been done on Intel Core 2,
y2 (t ) ~6320@1.86GHz
U A (t ) machine and a fair comparison of
computational burden is shown in Table 2 which
indicates that due to complexity, proposed algorithm
Figure 3. Schematic diagram of laboratory scale turbo-generator plant
consumes more time to tune the PID parameters.
7. Simulation Results Table2 Computational burden

Turbo- Generator plant which is nonlinear coupled Technique Computational time( sec)
MIMO system is used to illustrate the effectiveness of the Fixed PID 0.0290
proposed scheme. The Input-Output relation is as follows: DRNN based PID Tuning 0.0500
u (k-1) GA based PID Tuning 3.6020
y1(k) = 0.4y1(k-1)+ 1 2  0.2u13(k 1)  0.5u2 (k 1) (6)
1+u1 (k-1)
u2 (k-1)
y2 (k) = 0.2y2 (k-1)+  0.4u23 (k 1)  0.2u1(k 1) (7)
1+u22 (k-1)
The GA is used to tune for the optimal PID parameters
that minimize the fitness function (Eq. 5). Therefore the
parameter tuning task of the PID controller using GA can
be considered by selecting the three parameters kp, ki and
kd such that the response of plant will be as closest to
desire response. The GA parameters chosen for
simulation are shown in Table 1.
Table 1 Parameters of GA
GA property Value/ Method
Population Size 30
Maximum Number of 100 Figure 4 System response to sum of delayed step function
generations
Fitness Function Quadratic form as in Eq (5)
Selection Method Roulette wheel selection
Crossover Probability 0.90
Crossover Method Arithmetic Crossover
Mutation Probability 0.01
Initially many individual solutions are randomly

generated to form an initial population, allowing the
entire range of feasible solutions that results in stable
response (0<kp1<0.7, 0<ki1<0.9, 0<kd1<0.3, 0<kp2<1,
0<ki2<1, 0<kd2<1). Initial solution can be generated by
following expression
x = a+ (b-a)*rand (8)
where a, b are upper and lower limits of the parameter

value kp, ki and kd and rand is a random number between Figure 5 Controlled inputs
0 to 1. System response to sum of delayed step function
is shown in Figure 4 and a fair comparison among three
4
is not possible in realistic time frame i.e. online /

real time.
 The parameter optimization is global and has there
is no connection with initial population.
 With lower rise time and controlled input the
proposed scheme is capable to better rack the
delayed step function compare to other techniques.
 Controlling input effort is reduced and this offers a
major advantage.
9. Acknowledgements
Tarun Varshney is with the Department of Electronic
Instrumentation Engineering of Moradabad Institute of
Technology, Moradabad, India and currently pursuing
Figure 6 Optimization Process of fitness function
research at Motilal Nehru National Institute of
Technology, Allahabad, India. He wishes to acknowledge
the all support and facility of Moradabad Institute of
Technology for his work.
10. References:
[1] D.E. Goldberg, ‘Genetic algorithm in Search,
Optimization and Machine learning’, Addison-
Welsey Pulbilcation Co. Inc, 1989.
[2] Dr A.H Jones and P.B. de Moura Oliveira, ‘Genetic
Auto-Tuning of PID controller’, International
Conference on Genetic Algorithms in Engineering
Systems: Innovations and Applications, Sheffield,
UK, pp 141-145, 1995
[3] Takhiro Ota, Sigeru Omatu, ‘Tuning of the PID
Figure 7 Controller’s Parameter for first input
Control Gains by GA’, IEEE Conference on
Emerging Technologies and Factory Automation,
Kauai, Hawaii, pp 272-274, 1996
[4] Y. Mitsukura, T. Yamamoto and M. Kaneda, ‘A
Genetic Tuning Algorithm of PID parameters’,
IEEE International Conference On Systems, Man,
And Cybernetics , Orlando, Florida, USA, vol. 1, pp
923-928,1997.
[5] Yasue Mitsukura, Toru Yamamoto and Masahiro
Kaneda, ‘A design of self tuning PID controllers
using genetic algorithm’, Proceeding of American
control conference, San Diego, California, pp 1361-
1365, 1999.
[6] Thomas F. Junge, Heinz Unbehauen, ‘Recursive
Figure 8 Controller’s Parameter for Second input identification of a Turbo- Generator plant using
structurally adaptive neural network’, Proceedings
of IEEE International Conference on Industrial
8. Conclusions Technology, Goa, India, pp 572-577, 2000.
This paper demonstrates that how GA can be used to [7] T O. Mahony, C J Downing and K Fatla, ‘Genetic
optimize the parameters of PID controller for turbo- Algorithm for PID Parameter Optimization:
generator plant. The features of the proposed scheme are Minimizing Error Criteria’, Process Control and
summarized as follows: Instrumentation, University of Stracthclyde, pp 148-
 Optimized parameter can be obtained by optimizing 153, 2000.
the PID parameters controller with GA. [8] Maja Atanasijevi C-Kunc, Rihard Karba,
 It is important to select permissible upper and lower ‘Multivariable controller tuning by genetic
limits of parameters otherwise an optimal solution algorithms’, International Conference on Information
5
Technology Interfaces, Pula, Croatia, pp 337-342, [20] Daniel Carmona morles, Jorge E. Jimenez-
2000. Hornero, Francisco Vazquez and Fernando Morilla,
[9] J.M. Herrero, X. Blasco, M. Martínez, J.V. Salcedo, “ Education toolfor optimal controller tuning using
‘Optimal PID tuning with genetic algorithms for non evolutionary strageties”, IEEE transaction on
linear process models’, 15th Triennial World Education, vol. 55, No. 1. 2012.
Congress, Barcelona, Spain,2002
[10] Ian Griffin, ‘On line PID controller tuning using
genetic algorithm’, Dublin City University, 2003. Biographies
[11] Farag and H. Werner, “Structure selection and Tarun Varshney was born in India
tuning of multivariable PID controller for an in 1978. He received the B.E. degree
industrial benchmark problem”, IEE proceeding of in Electronics & Instrumentation
control theory and applications, vol.153, no 3, pp Engineering from Rajasthan
262 – 267, 2006. University, Jaipur, India, in 2001,
[12] Jih-Gau Juang, Ming-Te Huang, and Wen-Kai Liu, M.Tech. in Control &
“ PID control using presearch genetic algorithms for Instrumentation from Motilal Nehru
MIMO systems”, IEEE Transactions on systems, National Institute of Technology,
man, and cybernetics —Part C: applications and Allahabad, India, in 2007. Currently he is pursuing Ph.D.
reviews, vol. 38, No. 5, pp 716-727, 2008. at NIT, Allahabad. His research interests include
[13] Jin-Sung Kim, Jin- Hwan Kim, Ji-Mo Park, Sung- intelligent control systems, multi-variable systems
Man park and Won-Yong Choe, ‘Auto Tuning PID modeling and neural networks.
controller based on improved Genetic algorithm for
reverse osmosis plant’, World Academy of Science, Satya Sheel received his Bachelors
Engineering and Technology, pp 384-389, 2008. (’68) in Electrical, Masters (’76) in
[14] Zhu Supeng, Fu Wenxing , Yang Jun and Luo Control systems and Ph.D. in
Systems and Control Engineering
Jianjun, ‘Applying Genetic algorithm to (’82). Currently he is a Professor at
optimization parameters of missile control system’, the National Institute of Technology
International conference on Hybrid Intelligent at Allahabad, India. He is a senior
system, Shenyang, China, pp 416-419, 2009. member of IEEE, fellow of
[15] Neenu Thomas, Dr. P. Poongodi, ‘Position Control Institution of Engineers (I) and has been India
of DC Motor using genetic algorithm based PID representative of AMSE (France). He is a reviewer for
journals, conferences and books in control,
controller’, Proceeding of world congress on
instrumentation. His current interests are intelligent
Engineering,Vol. II, London, UK, 2009. control of multivariable systems, PID controller tuning,
[16] P. Poongodi, S. Victor, ‘Genetic Algorithm based embedded systems, process control, machinery health
PID controller design for LTI system via reduced diagnostics and control engineering education.
order model’ International conference on
Instrumentation, Control and Automation,
Bandung, Indonesia, pp 115-118. 2009.
[17] Mohammed Obaid Ali, S. P Koh, K. H. Chong,
S.K. Tiong and Zeyad Assi Obaid, ‘Genetic
Algorithm Tuning based PID controller for liquid
level tank system’, International Conference on
Man-Machine System, Batu Ferringhi, Penang,
Malaysia, pp 4A5-1- 4A5-5, 2009.
[18] Monica Patrascu, Adrian Bogdan Hanchevici, Ioan
Dumitrache , ‘Tuning of PID Controllers for Non-
Linear MIMO Systems Using Genetic Algorithms’,
18th IFAC World Congress Milano (Italy), pp
12644-12649, 2011.
[19] Tarun Varshney, Satya Sheel, “A New Online
Tuning Approach for PID Control of Multivariable
Systems using Diagonal Recurrent Neural
Network”, IEEE International Conference on
Control System, Computing and Engineering,
Penang, Malaysia,2011.
6
Closed Loop Controller for Stochastic Systems with Uncertain Parameters
Mohamed F. Hassan and Hala A. Mourad

Electrical Engineering Department, College of Engineering,
Kuwait University, 13060 Kuwait
phone: +965-2498-7634; fax: +965-2481-7451
m.f.hassan47@gmail.com & m.fahim@ku.edu.kw
Abstract state estimation of linear discrete systems with uncertain

In this paper an observer-based stochastic controller is parameters, interested readers can refer to [3] and
developed for linear stochastic discrete-time dynamical references therein. On the other hand, for nonlinear
systems with uncertain parameters. A state estimator, dynamical systems, the stability, and hence the
based on Kalman filter, is firstly developed which robustness of the estimator become worse. Several
incorporates in its structure the statistical information of algorithms have been proposed to estimate the states of
the uncertain parameters. Although the developed nonlinear systems [4, 5, 6, 7, and 8] and references
estimator uses the nominal values of the system therein. In general, the extended Kalman filter (EKF) is
parameters; it showed to be stable, robust and gives the most extensively used algorithm for state estimation
satisfactory results. Such an estimator is then used to of nonlinear systems. However, it is not the optimal
design an observer-based controller for discrete-time estimator, and it may diverge if the system acquires
linear stochastic dynamical systems. The effectiveness of strong nonlinearities, or the initial estimate of the state is
the theoretical derivation of the developed estimator and wrong, or the process is modeled incorrectly, owing to
the observer-based controller are illustrated via a linearization [9, 10, 11].
numerical example and a practical case study. In this paper, a new approach is proposed to estimate the
states of linear stochastic discrete-time dynamical system
Keywords: Discrete-time systems, linear systems, with uncertain parameters. The system model and the
nonlinear systems, state estimation, stochastic systems, measurements are assumed to be corrupted by
control design. uncorrelated zero mean white Gaussian noise sequences.
The parameters of the system are assumed to be uncertain,
1. Introduction which obviously leads to an estimation problem of
Since most, if not all, real control systems are stochastic nonlinear (bilinear) stochastic discrete-time dynamical
by their nature, it has been essential to develop and system. The new filter handles this problem and its
enhance stochastic control theory which mainly deals performance is compared with that of the KF, when the
with this type of problems. With the progress of parameters are assumed to be certain, and with the EKF,
computational tools, and the development of state space when the parameters are decided to be estimated. The
approach, it became a systematic procedure to represent proposed estimation technique is used to design an
system models by state equations rather than frequency observer-based controller for linear stochastic discrete-
domain models. In practice, the states of the system are time system with uncertain parameters. For linear
not accessible to direct measurements. Therefore, it was stochastic systems with certain parameters, the separation
mandatory to develop a theory for state reconstruct of the principle holds. Therefore, controller design and state
system in order to generate the desired control strategies. estimator can be performed separately. This leads to an
For deterministic linear finite–dimensional time invariant optimal closed loop stochastic controller [12]. However,
systems, an observer was first introduced by Luenberger in case of stochastic systems with uncertain parameters,
[1] leading to the asymptotic estimation of the state. In the separation principle does not hold, since additional
case of linear dynamical system with random terms will appear in the state space representation of the
disturbances, where the stochastic phenomena appears, augmented system due to the uncertain components of
the state estimation problem is solved using Kalman filter the system parameters. Although this is the case, we can
(KF) [2]. In standard Kalman filter, all the system still design a closed loop suboptimal observer-based
characteristics (i.e. system model, initial conditions, and controller provided that it leads to an asymptotically
noise characteristics) have to be specified a priori. stable system. To show the effectiveness of the
However, if there is an uncertainty in any of these developed estimation approach, illustrative examples are
characteristics, the filter may not be robust enough. For solved using the proposed estimator, the KF and the EKF
7
with parameter estimation. MATLAB is used as a bk  [a11k , a12k ,..., a21k , a22k ,...]T ,
programming tool and Monte Carlo simulation is applied (2)
to compare the behavior of the different filters. An ck 1  [c11k 1 , c12k 1 ,..., c21k 1 , c22k 1 ,...]T
observer-based controller, using the developed estimator,
In the above model it is assumed that
is designed to regulate a DC motor. The obtained results
x , bk , ck 1 , wk , vk 1 are all independent. Moreover, it is
show the applicability and efficiency of the proposed
procedure in handling this problem. assumed that the model has the following properties:
The rest of the paper is organized as follows. In section 2, E{ xk wTj }  0,  k  j ; E{ xk v Tj }  0,  k , j ;
the control problem for stochastic linear discrete-time
dynamical systems with uncertain parameters is E{ yk vTj }  0,  k  j ; E{ xk bTj }  0,  k  j ;
formulated. Section 3 is devoted to the presentation of the (3)
E{ xk cTj }  0,  k , j ; E{ yk cTj }  0,  k  j ;
mathematical structure of the developed estimator. The
dynamics of the observer-based controller is presented in E{bk cTj }  0,  k , j
section 4. The algorithm used to implement the closed
Since the parameters of the system are uncertain, then
loop stochastic controller is given in section 5. In section
we have a bilinear estimation problem rather than linear,
6, illustrative examples are solved to show the
which intern leads to a non-Gaussian estimation problem.
effectiveness of the proposed design approach. The paper
Our objective is to get a good estimator for the state
is concluded in section 7.
vector, and hence use the estimated states to generate the
desired control signal.
2. Problem Formulation
Let us define the following stochastic control problem:
 1 k f 1 
3. The Developed Estimator
  2 

2 Let us for the moment consider the following problem:
min J  E   xk  x d  uk N   (1-a)
 2 k  k  M  x k 1  Ak x k  w k (1'-a)

Subject to: yk 1  C k 1 xk 1  vk 1 (1'-b)
x k 1  Ak x k  B k u k  d  w k (1-b) The system model (1') has the same definitions and
properties of the system model (1) in section 2.
yk 1  C k 1 xk 1  vk 1 (1-c)
State estimation problems with uncertain parameters is
where: xk  R n is the state vector, yk  R m is the usually solved by treating the parameters as static states,
output vector, u k  R r
is the control vector, d  R n is i.e.  k 1   k with the prior at k=0 being N (  ,  ) ,
a constant vector, x d  R n is the desired state vector, where  is the vector of the mean or nominal values of
the parameters and  is its covariance matrix. This
Ak  aijk   R nxn is the system matrix with uncertain
approach increases the dimensionality of the problem,
parameters, aijk ; i , j  {1, 2,...., n} assumed to be and hence the computational burden. To avoid this
difficulty, the nominal values of the parameters will be
uncorrelated white Gaussian random variables with the
used in the estimator. However, the impact of this
following statistics aijk   (aij ,  a2 ) , Bk  R nxr is the approximation on the dynamics of the estimator will be
ij
investigated.
control transition matrix, wk  R n is a zero mean Assume that the filtered estimate xˆ k|k and its associated
white Gaussian input noise vector with covariance matrix
covariance matrix Pk|k are given at the k th sampling
Qk  E{wk wkT }  R nxn , C k 1   cijk 1   R mxn is the
 
instant, and used to approximate the conditional
output measurement matrix with uncertain
parameters, cijk 1 ; i  {1, 2,...., m}, j  {1, 2,...., n} assumed probability density function p( xk | Y k ) by Gaussian
distribution with conditional mean and covariance matrix
to be uncorrelated white Gaussian random variables with
as given above.
the following statistics cijk 1   (cij ,  c2ij ) , vk 1  R m is The filtered estimate of system (1') and its associated
a zero mean white Gaussian output noise vector with covariance matrix are given by:
a) The predicted estimate of the state vector xˆ k 1|k , its
covariance matrix Rk 1  E{vk 1vkT1}  R mxm
associated covariance matrix Pk 1|k , and the predicted
M  R nxn is positive semi-definite weighting matrix for
the states, N  Rrxr is positive definite weighting matrix estimate of the output vector yˆ k 1|k , are given by:
for the control, x  R n is the initial conditions of the xˆ k 1|k  Axˆ k |k (4)
states assumed zero mean random Gaussian vector with Pk 1|k  APk |k AT  ˆ k |k  kˆ kT|k   k  Qk (5)
covariance matrix P  P |  E{ x xT }  R nxn , and
yˆ k 1|k  Cxˆ k 1|k (6)
finally k  {0,1, 2,...} is the discrete time instant.
2
b) The filtered estimate of the state vector xˆ k 1|k 1 , its
Let bk  R n , ck 1  R nm be the parameter vectors
associated covariance matrix Pk 1|k 1 , and the filtered
defined as:
estimate of the output vector yˆ k 1|k 1 , are such that:
8
xˆ k 1|k 1  xˆ k 1|k  K k 1 y k 1|k (7) Equation ( 1' -b) can also be written in the form:
Kk1  Pk1|kCT [CPk1|kCT ˆk1|kUk1ˆkT1|k  Lk1  Rk1 ]1 yk 1  (C  C k 1 )( xˆ k 1|k  x k 1|k )  vk 1
(8) yk 1  Cxˆ k 1|k  Cx k 1|k  C k 1 xˆ k 1|k  C k 1 x k 1|k  vk 1

Pk 1|k 1  ( I  K k 1C ) Pk 1|k (9) (20)
where C is the matrix contains the nominal (mean)
yˆ k 1|k 1  Cxˆ k 1|k 1 (10)
values of the parameters, while C k 1 contains the
Proof:
a) Derivation of the Prediction Estimator and the perturbations around these nominal values.
Corresponding Covariance Matrix: Let ck 1  [c11 , c12 ,..., c21 ,..]T  R nm be the vector
k 1 k 1 k 1
Equation ( 1' -a) can be rewritten as: contains the deviation of the parameters of the matrix
xk 1  ( A  A k )( xˆ k |k  x k |k )  wk C k 1 around their nominal values C . Then, by
(11)
xk 1  Axˆ k |k  Ax k |k  A k xˆ k |k  A k x k |k  wk rearranging (20), one gets:
where A is the matrix contains the nominal (mean) yk 1  Cxˆ k 1|k  Cx k 1|k ˆ k 1|k ck 1  k 1|k ck 1  vk 1
values of the parameters, while A k contains the (21)
where:
perturbations around these nominal values.
 xˆ kT1|k O O........... 
Let b  [a , a ,..., a , a ,...]T  R n be a vector
2
k 11k 12k 21k 22k  

O xˆ kT1|k O........... 
contains the deviation of the parameters of the matrices ˆ k 1|k    R mx ( nm ) ,
Ak around their nominal values A , and define: .................................. 
 T 
 xˆ kT|k O O...........   O O ............... xˆ k 1|k 
  (22)
O xˆ kT|k O...........   x kT1|k O O........... 
ˆ k |k
2
   R nxn ,  
..................................  O  kT1|k O........... 
x
   k 1|k     R nx ( nm )
 O O ............... xˆ kT|k  .................................. 
(12)  T 
 x kT|k O O...........   O O ............... x k 1|k 
 
O x kT|k O...........  From (21), the predicted output is given by:
 k |k
2
   R nxn yˆ k 1|k  Cxˆ k 1|k (23)
.................................. 
  Accordingly, y k 1|k is such that:
O O ............... x kT|k 
y k 1|k  Cx k 1|k ˆ k 1|k ck 1  k 1|k ck 1  v k 1 (24)
where O  [0 0 0.........]  R n is a vector of zero
Hence:
elements.
Then, (11) can be rearranged in the following form: Py k1|k y k 1|k  E{ y k 1|k y k 1|k }
(25)
xk 1  Axˆ k |k  Ax k |k  ˆ k |k bk   k |k bk  w k (13)  CPk 1|k C T ˆ k 1|k Uk 1ˆ kT1|k  Lk 1  Rk 1
The predicted value of the state vector is given by: where:
xˆ k 1|k  E { x k 1 | Y k } (14) U k 1  diag[ c211  c2 ................... c2mn ]
k 1 12k 1 k 1
Using the model properties (3), the predicted estimator is Lk 1  diag[l11k 1 l22k 1 .....................lmmk 1 ] (26)
such that:
n
xˆ k 1|k  Axˆ k |k (15)
lk 1ii   Pk 1|k jj  c2ij
k 1
From (13), (15), the prediction error is given by: j 1
x k 1|k  Ax k |k  ˆ k |k bk   k |k bk  wk (16) b) Derivation of the Filtered Estimator and the
Corresponding Covariance Matrix:
The covariance matrix of the estimation error is such that:
The filtered estimate of xk 1 is given by:
Pk 1|k  E{ x k 1|k x kT1|k } (17)
xˆ k 1|k 1  E{ xk 1 | Y k , yk 1}
Using the model properties (3), Pk 1|k is given by: (27)
 xˆ k 1|k  Pxk 1 y k 1|k Pyk11|k y k 1|k y k 1|k
Pk 1|k  APk |k AT  ˆ k |k  kˆ kT|k   k  Qk (18)
where: Pxk 1 y k 1|k  E{ xk 1 y kT1|k }
(28)
 k  diag[ a2 11k
 a2 12k
................... a2nn ] (19-a)  E{( xˆ k 1|k  x k 1|k ) y kT1|k }
k
 k  diag[e11 k
e22k .....................ennk ] (19-b) Using (24) in (28), and due to the independency
n between x k 1|k , ck 1 , vk 1 , one gets:
eiik   Pk |k jj  a2ij (19-c)
j 1
k Pxk 1 y k 1|k  Pk 1|k C T (29)
9
Let K k 1  Pxk 1 y k 1|k Pyk11|k y k 1|k (30) d) The covariance matrix Pk 1|k , and the predicted
estimate of the output yˆ k 1|k are as given by (5), (6).
Using (25), (29) in (30) one can get K k 1 as given by
(8). Hence: e) The filtered estimate of the state vector xˆ k 1|k 1 , its
xˆ k 1|k 1  xˆ k 1|k  K k 1 y k 1|k (31) associated covariance matrix Pk 1|k 1 , and the
The filter estimation error is such that: filtered estimate of the output vector yˆ k 1|k 1 are as
x k 1|k 1  x k 1|k  K k 1 y k 1|k (32) given by (7)-(10).
Using (32), the covariance matrix of the filter estimation The flowchart of the observer-based controller is
error is such that: given in the following section.
Pk 1|k 1  ( I  K k 1C ) Pk 1|k (33)
5. The Flowchart of the Algorithm
4. Control Design
In the system model (1), since parameters of the system
as well as the states are random, the system is bilinear START
rather than linear. Therefore, the separation principle
does not hold. However, the problem will be treated as
deterministic and the resulted control strategy will be INPUT
applied to the stochastic problem. It is obvious in this
case that the resulting closed loop control system is
suboptimal rather than optimal.
Let us consider the following servo-mechanism discrete-
time linear deterministic system:
k 1
1 f  

2
min J  xk  x d  uk
2
(34) CALCULATE

2 k k  M N 

Subject to:
x k 1  Ax k  B k u k  d (35)
n
where x k  R is the state vector assumed to be
deterministic, A  R nxn is the matrix contains the
nominal values of the parameters, while other variables
are as defined in section 2. No
By writing the Hamiltonian of this problem, and from the If
necessary conditions of optimality, the expression for the
control is such that [13]:
uk   N 1 BkT AT   Prk  M  xk   k  Mx d  Yes
1
(36)
T 1 END
where Prk  M  A Prk 1  A with Prk  [0] (37)
f
with   I  B k N 1B kT Prk 1

(38) 6. Simulation Results
6.1 Illustrative Example :
k   AT Prk 1 1Bk N 1BkT k 1  Mx d  AT Prk 1 1d To investigate the effectiveness of the developed
 AT k 1 with k f  0 (39) estimator, we employ firstly the following numerical
values of a three dimension (3-D) discrete-time linear
Since the state vector xk is not available for direct system model:
measurement, the estimate of this vector xˆ k |k will be  1 0.009983 0 
used in conjunction with the state feedback controller A   0.009889 0.9949 0.0091  , C  [1 0 0] ,

(36). Therefore, the dynamics of the observer-based   1.98  0.9874 0.8305 
closed loop stochastic controller at the k th iteration can
be summarized as follows: x  [1 1 1]T , xˆ |  [0 0 0]T ,
a) Calculate, off-line, Prk ,  k for k  k ,...., k f 1 using P |  I where I  R3 x 3 is the unity matrix.
(37)-(39). The states of the system are estimated using the
b) Calculate the control signal using (36). developed filter, KF, and EKF. Different values of
c) Calculate the predicted value of the state vector as Q, R and different levels of uncertainties in the system
given by:
parameters are used in our simulation. For comparison
xˆ k 1|k  xˆk |k  B k u k (40)
purposes, Monte Carlo simulation is applied to explore
the qualitative behavior of the root mean square errors
10
(RMSE) of the estimated states resulting from the

application of the three filters. This is defined by: Modified Kalman Filter
4

MI 2
(k )  xˆ j (k | k ))
RMSE State 2
i 1
(x j 3
RMSE (x j (k ), xˆ j (k | k ))  2
MI 1
0 100 200 300 400 500 600 700 800 900 1000
Kalman Filter
for k  1,..., k f and j  1,...., n (41)

4
RMSE State 2
3
where MI is the number of Monte Carlo iterations. 2
Moreover, the root mean square error index (RMSE 1

0 100 200 300 400 500 600 700 800 900 1000
index) for each state is calculated using the following 2

x 10
147 Extended Kalman Filter
formula:
RMSE State 2
1.5
 
kf MI 2
1
k 0
(x
i 1 j
(k )  xˆ j (k | k )) 0.5
RMSE index (x j )  (42) 0
(k f )(MI )
0 100 200 300 400 500 600 700 800 900 1000
time (sec)
Table (1) shows the RMSE indices calculated over 100 Figure 2: RMSE of x 2 (Case-a).
Monte Carlo iterations for each case study while using
the three filters. Modified Kalman Filter
30
RMSE State 3
Table 1: RMSE index (MI = 100) 20
10
Case xi Q R Pn% RMSE index 0

0 100 200 300 400 500 600 700 800 900 1000
MKF KF EKF
Kalman Filter
30
RMSE State 3
a 0.1 0.1 10 0.2597 0.269 20
10
b 0.5 1.0 20 0.7947 1.821

1 Div.
0
0 100 200 300 400 500 600 700 800 900 1000
c 0.5 1.0 30 0.8235 8.233 4

x 10
150 Extended Kalman Filter
RMSE State 3
d 1.0 1.0 30 0.8717 26.444 3
a 0.1 0.1 10 1.8834 1.887 1
0
0 100 200 300 400 500 600 700 800 900 1000
b 0.5 1.0 20 19.687 25.099 time (sec)
2 Div.
c 0.5 1.0 30 61.726 66.207 Figure 3: RMSE of x3 (Case-a).
d 1.0 1.0 30 576.66 686.21
a 0.1 0.1 10 10.054 10.09 1
Modified Kalman Filter
b 0.5 1.0 20 95.214 125.57

RMSE State 1
0.9
3 Div. 0.8
c 0.5 1.0 30 299.092 345.97 0.7
0 100 200 300 400 500 600 700 800 900 1000
d 1.0 1.0 30 2.23e+3 2.87e+3 Kalman Filter
15
RMSE State 1
10
From Table 1, the proposed filter has the best 5
performance in all cases. It has the minimum RMSE 0

0 100 200 300 400 500 600 700 800 900 1000
index calculated over 100 Monte Carlo iterations, 1.4
Extended Kalman Filter
whereas the (EKF) fails to estimate the states since it

RMSE State 1
1.2
diverges. The RMSE plots of the states using the three 1
filters for Case-a are shown in Figures (1-3), and for

0.8
0 100 200 300 400 500 600 700 800 900 1000
Case-b are shown in Figures (4-6). time (sec)
1
Figure 4: RMSE of x1 (Case-b).
RMSE State 1
0.5
200
RMSE State 2
0 150
0 100 200 300 400 500 600 700 800 900 1000
100
50
Kalman Filter
1 0
0 100 200 300 400 500 600 700 800 900 1000
RMSE State 1
Kalman Filter
300
0.5
RMSE State 2
200
0 100
0 100 200 300 400 500 600 700 800 900 1000
0
0 100 200 300 400 500 600 700 800 900 1000
Extended Kalman Filter 149
x 10 Extended Kalman Filter
1 8
RMSE State 1
RMSE State 2
0.5 4
0
0 0 100 200 300 400 500 600 700 800 900 1000
0 100 200 300 400 500 600 700 800 900 1000 time (sec)
time (sec)
Figure 1: RMSE of x1 (Case-a). Figure 5: RMSE of x 2 (Case-b).
11
Modified Kalman Filter k 1  1 0.005 0  k   0 

600
        
   0 0.9994 0.0749      0 150 (46)
RMSE State 3
400
k 1 k
      
iak 1  0  0.9632  0.9238 iak  0.5023
200
k 1 
0 100 200 300 400 500 600 700 800 900 1000
 
Kalman Filter
yk 1  [0 1 0] k 1 
1000
(47)
RMSE State 3
500 i 
 ak 1 
0
0 100 200 300 400 500 600 700 800 900 1000
where, the system model is corrupted by an input noise
x 10
149 Extended Kalman Filter w (k )  N (0, 0.1) and measurement disturbance
10
v (k )  N (0, 0.01) .
RMSE State 3
The system parameters KT , Kb , La and R ' , are

5
0
0 100 200 300 400 500 600 700 800 900 1000
assumed to be uncertain with a standard deviation equals
time (sec)
to certain percentage Pn of their nominal values.
Figure 6: RMSE of x3 (case-b) x  [10 80 0.5]T , xˆ |  [5 79 1]T , P |  I  R3 x 3
The RMSE indices of the states calculated over 100
6.2 Case Study: Monte Carlo iterations using the three filters are shown in
The transfer function of a DC motor is given by [14]: Table 2. RMSE plots for Case-b are shown in Figures (7-
 (s) KT 9).
 (43)
E ( s ) s[( sLa  R ' )( sJ  B )  KT K b ]
Table 2: RMSE index, Q=0.1, R=0.01 (MI = 100).
where, Case State Pn% RMSE index
 is the motor angle displacement.
d MKF KF EKF
 is the motor rotating speed, where   .
dt a 5 8.6057 8.6057
ia is the armature winding current. 
b 10 9.3070 9.3070
E is the applied voltage.
KT is the motor torque constant. a 5 0.1034 0.1231
b  Div.
Kb is the back emf constant. 10 0.1044 0.2213
La is the armature winding inductance. a 5 6.6085 12.4121
ia
'
R is the armature winding resistance. b 10 15.5597 34.4408
J is the moment of inertia of rotor and load.
B is the damping coefficient.
The state vector x is defined by x T  [  i a ]T . 12
For the above model, it is assumed that the rotating

RM SE State 1
10
speed of the motor  is measured at the output. 8
6.2.a) State estimation under no control: 6
The continuous state equation of the DC motor is: 4

0 100 200 300 400 500 600 700 800 900 1000
   0 1 0     0  Kalman Filter
d    
KT / J      0  E (44)
12
   0 B/J
RM SE State 1
10
dt
ia   0  K b / La  R / La  ia  1/ La 
     8
4
  0 100 200 300 400 500 600 700 800 900 1000
  Extended Kalman Filter

y  [0 1 0]   (45) 8
ia 
RM SE State 1
7
 
where E=150 is the reference input. 6
The above model is discretized using Euler’s method 5

0 100 200 300 400 500 600 700 800 900 1000
with sampling period Ts= 0.005. Therefore, the discrete time (sec)
model is such as:
Figure 7: RMSE of  (case-b)
12

1
RMSE State 2
0.5
0
0 100 200 300 400 500 600 700 800 900 1000
Kalman Filter
1
RMSE State 2
0.5
0
0 100 200 300 400 500 600 700 800 900 1000
Extended Kalman Filter

1
RMSE State 2
0.5
0
0 100 200 300 400 500 600 700 800 900 1000
Figure 11: Actual and Estimated motor angle

time (sec)
Figure 8: RMSE of (Case-b). displacement .

60
RMSE State 3
40
20
0
0 100 200 300 400 500 600 700 800 900 1000
Kalman Filter
150
RMSE State 3
100
50
0
0 100 200 300 400 500 600 700 800 900 1000
150
x 10 Extended Kalman Filter
3
RMSE State 3
0
0 100 200 300 400 500 600 700 800 900 1000
time (sec)
Figure 9: RMSE of (Case-b).

Figure 12: Actual and Estimated motor
From the simulation results, the proposed estimation rotating speed .
procedure showed to be the best, and leads to the
minimum value of the RMSE index. Therefore, it will be
used in the implementation of the observer-based
controller as shown in Figure (10).
+ y
r System Measurement
+
u
Controller MKF
Figure 10: Block diagram of the closed loop system.
6.2.b) Observer-based Controller:

For the problem in hand, it is desired to regulate the
system to the rotating speed  d  76.8085 rpm . The
Figure 13: Actual and Estimated armature
initial values for the states, the estimated states, and winding current .
covariance matrix P are as follows:
T T 3 x3
x  [5 40 1] , xˆ |  [0 0 0] , P |  I  R
The state weighting matrix is M  diag[0 1 1] , the
control weighting matrix is N =[1], the covariance matrix
of the input noise is Q  0.1I , the covariance matrix of
the output noise is R  0.01 , and the percentage change
in the parameters is Pn  10% . The actual and estimated
closed loop trajectories as well as the control signal are
shown in Figures (11-14).
Figure 14: Feedback Control u.
13
7. Conclusion and System Engineering ACSE, vol. 7, no.2, pp1-

In this paper, an observer-based controller is developed 8, Nov. 2007.
for linear stochastic discrete-time dynamical systems [11] S.Abraham Lincon and D.Sivakumar and J.Prakash,
with uncertain parameters. A state estimator, based on State and Fault Parameter Estimation Applied To
Kalman filter, is firstly developed which incorporates the Three-Tank Bench Mark Relying On Augmented
statistical information of the uncertain parameters in its State Kalman Filter. Automatic Control and
structure. Although this problem is treated in the System Engineering ACSE, vol.7, no.1, pp 33-41,
literature as a nonlinear estimation problem; with the May 2007.
developed approach the system is still treated as linear [12] J.S. Meditch, Stochastic Optimal LinearEstimation
and without any increase in its dimensionality. As a result, and Control. McGraw Hill, 1965.
we have less computational burden and better numerical [13] D.E. Kirk, Optimal Control Theory: An Introduction,
properties. Based on the developed filter, an observer- Englewood Cliffs, N.J., Prentice Hall, 1970.
based feedback controller is proposed. Simulation of [14] S.K. Yang , An experiment of state estimation for
different illustrative examples showed that the state predictive maintenance using Kalamn filter on a
estimation using the proposed filter is better than KF, DC motor. Reliability Engineering and System
while the EKF diverged and failed to estimate the system Safety, 2002.
states. Moreover, although the separation principle does
not hold, simulation results showed that the regulated Biographies
system is asymptotically stable, and the states of the DC
Mohamed F. Hassan was born in
motor reached the desired ones. However, it is still Cairo, Egypt on June 21, 1947. He
necessary to analyze the stability of the controlled system
received the B.Sc. and M.Sc.
more rigorously in a theoretical frame work. This degrees in Electrical Engineering
problem will be treated in the future. from Cairo University, Faculty of
Engineering, Cairo, Egypt in 1970
8. References and 1973 respectively, and D.Sc.
[1] D.G. Luenberger, Obserevers for multivariable (Doctor d’Etat) in Systems Science and Automatic
systems. IEEE Trans. Automat Control AC-11, Control from Paul Sabatier University, Toulouse, France,
pp.190-7, 1966. 1978. He is a Professor of Systems Science and Control
[2] R.E. Kalman R.E. , Anew approach to linear filtering Engineering in the Electrical, College of Engineering and
and prediction problems. Transactions of ASME- Petroleum, Kuwait University since 2002 till now. He has
Journal of Basic engineering, 82-D, pp. 35-45, been a faculty member in Electronics and
1960. Communication Department, Faculty of Engineering,
[3] T. Cimen State-dependent Riccati equation (SDRE): Cairo University since 1978 where he held the positions
A survey. Proceeding of the 17th World Congress, of Assistant Professor associate Professor and Professor .
The International Federation of Automatic Control, His research activities are in the areas of large scale
Seoul, Korea, July 6-11, pp.3761-3774, 2008. systems, optimization theory, stochastic control theory,
[4] E.L. Haseltine and J.B. Rawlings, Critical evaluation estimation theory and adaptive control. Prof Hassan is a
of extended Kalman filtering and moving horizon co-author of one book, author of many book chapters and
estimation. Ind.Eng.Chem. Res.,vol. 44, no. 8, encyclopedia chapters . Moreover, he is an author and co-
2005 author of more than 150 technical papers in most of the
[5] M.S. Arulampalam, S. Maskell, N. Gordon, and T. international journals and conferences. Prof. Hassan is a
Clapp A., Tutorial on Particle Filters for Online reviewer of many international journals and conferences.
Nonlinear/Non-Gaussian Bayesian Tracking. IEEE
Trans.On Signal Processing, vol. 50, no.2, pp174-
Hala A. Mourad was born in
188, Feb. 2002.
Kuwait on January 11, 1966. She
[6] G. Evensen , Data Assimilation: The Ensemble
received the B.Sc. and M.Sc.
Kalman Filter. Springer-Berli, 2009.
degrees in Electrical Engineering
[7] E.A. Wan and R. var der Merwe, The unscented
from Kuwait University, College of
Kalman filter for nonlinear estimation. Proc. IEEE
Engineering and Petroleum, Kuwait
Symposium on Adaptive Systems for Signal
in 1987 and 2011 respectively.
Processing, Comm. And Control, Lake Louise,
Canada, pp. 153-158, Oct. 2000.
[8] D. Simon, Optimal State Estimation: Kalman, H
infinity and Nonlinear approach.Wiley, 2006.
[9] Y. Govindarajan and D. Sivakumar , Attack Analysis
of Robust Watermarking Technique for Patent
Infringement.Graphics,Vision and Image
processing GVIP, vol. 11, no.3, pp39-45, Jul. 2011.
[10] A. Asokan and D. Sivakumar, Fault Detection and
Diagnosis for a Three-tank system using
Structured Residual Approach. Automatic Control
14
A Novel Nine-Level MPC Inverter for Direct Torque Control Induction Motor
Drive
O. Chandra Sekhar, K.Chandra Sekhar

Dept. of Electrical and Electronics Engg, Vignan`s Lara Institute of Technology and Science, Vadlamudi, Guntur, India
Dept.of Electrical and Electronics Engg, R.V.R & J.C College of Engg, Chowdavaram, Guntur, India
Sekhar.obbu@gmail.com,cskoritala@gmail.com
Abstract connection of power semiconductors are avoided and a

In this paper a novel nine–level inverter system for Direct multilevel output voltage is also possible.
Torque Control (DTC) induction motor drive is presented.
DTC strategy of a nine–level Multi Point Clamped (MPC) Feeding electrical motors with multilevel inverters can
Voltage Source Inverter (VSI) fed induction motor is greatly improve the motor performance. The harmonic
implemented. Nine-level MPC inverter is an advanced content of the voltage applied to the motor’s terminals
technology and are commonly applied to high power will be greatly reduced. Recently, the high dv/dt
electrical drives instead of the two-level VSI.Nine-level produced by modern power electronic based inverters has
inverter has 729 space vector switching states and there been identified as the source of several motor failures [8],
are 106 effective vectors are possible. The proposed nine- [9].The ability of multilevel inverters to deliver a
level inverter scheme is capable for enough degrees of multistepped voltage waveform will prevent motor
freedom to control both electromagnetic torque and stator failures by reducing the voltage change rate of the
flux with very low ripple. From the simulation results applied voltage. Any of the three prominent multilevel
shows that feeding electrical drive with nine-level topologies can be used in motor drive applications.
inverter can greatly improves the drive performance as
compared to the seven-level inverter. The performance of MPC is so called since in their architecture there are
this control method has been demonstrated by several points clamped to specific voltages using some
simulations performed using a versatile simulation components. Even diode-clamped converters belong to
package, Matlab/Simulink. this family because the bus between two switches is
clamped by a clamping diode. Furthermore, when the
number of voltage level is odd, the converters are called
Neutral Point Clamped (NPC) because the neutral point
Keywords: DTC, Nine-level inverter Multi point is clamped, has found wide application in medium-
Clamping, Switching table, Induction motor, voltage high-power applications [10][11]. The main
Matlab/Simulink. features of the MPC converter include reduced dv/dt and
Total Harmonic Distortion (THD) in its AC output
1. Introduction voltages in comparison to the conventional two level
Multilevel converters have recently increased interest in converters. As in any multilevel converter it can be used
the research and industry communities [3]–[7]. In these in the medium-voltage applications to reach a certain
kind of converters, the output voltage can take several voltage level without series connection of power
discrete levels of equal magnitude. The multilevel semiconductors.
converter first proposed in [1] was aimed at reducing the
harmonic content of generated voltage or current In principle, DTC method is based on instantaneous
waveforms. The harmonic content of such a waveform is space vector theory. By optimal selection of the space
greatly reduced, if compared with a two-level waveform. voltage vectors in each sampling period, DTC achieves
Nowadays, there are three prominent multilevel effective control of the electromagnetic torque and the
topologies. Multilevel structures derived from those stator flux on the basis of the errors between theirs
proposed in [2] and [3] are known as diode-clamped (or references and estimated values. It is possible to directly
neutral-point-clamped) and flying-capacitor multilevel control the inverter states through a switching table, in
converters, respectively. Both of them are based on a order to reduce the torque and flux errors within the
serial connection of low-voltage power semiconductors. desired bands limits[12][13]. The present work is based
The use of voltage clamps to equally share the input on the study of the application of DTC to the nine-level
voltage between the serial-connected devices and at the MPC VSI, and the advantages that can be obtained from
same time having a multilevel output voltage is a using this topology when compared to the seven-level
common feature of these two structures. Thus, by this inverter.
means, the known problems posed by the serial
15
The paper is organized as follows: the concept of nine - are cut in 1/8th with reduced harmonics of output
level MPC VSI is discussed in section 2 and waveforms for the same switching frequency, and power
representation of proposed DTC scheme is discussed in rating is increases.
section 3. The simulation results and discussion of the Relatively to the 7-level inverter which is only capable to
proposed method are exposed at the next section. Finally produce 343 space vector voltages, a 9-level inverter has
conclusions are given in the last section 93=729 switching states are possible as shown in figure.
If voltages of eight capacitors are equivalent, some
2. Nine-Level MPC VSI and Related Output switching vectors are overlapped and there are 106
Voltage Vectors effective vectors.
The architecture of MPC is several points clamped to According to the magnitude of the voltage vectors, we
specific voltages using some components. Even diode- divided them into thirteen different groups:
clamped converters belong to this family because the bus
between two switches is clamped by a clamping diode.
Furthermore, when the number of voltage level is odd,
the converters are called Neutral Point Clamped (NPC)
because the neutral point is clamped. As Figure.1 shows,
the 9-level leg is completely different: in MPCs the
voltages are clamped using couples switch-diode instead
of using a simple diode. Anyway, given a number of
levels, the number of switches needed by MPC is the
same needed by diode-clamped. The control of MPC leg
is more complicated in the respect of other topologies.
Even this kind of converter allows finding
complementary couples of switches. The constrain so
introduced is not physiologically necessary, but it is a
simple way to simplify the control scheme and the
switching table (Table I).
TABLE I
SWITCHING STATES OF A NINE-LEVEL
INVERTER
Figure.1: Nine-level multi point Clamped Inverter leg

TABLE II
Magnitudes of the voltage vectors
Group Magnitudes of the voltage vectors

1 [V0]
2 [V1, V2, V3, V4, V5, V6]
3 [V44, V45, V46, V47, V48, V49]
Furthermore, to avoid shortcut, T5 and T6 must be 4 [V123, V124, V125, V126, V127, V128]
complementary controlled. The same must happens for 5 [V135, V137, V139,V141,V143, V145],
T′5 and T′6. Table I is a possible switching table for a 9- [V180,V181,V182,V183,V184,V185,V186,
6
level MPC leg; there is an intra-phase redundancy only V188,V190,V192,V194,V196]
for the middle level. Anyway, it is better to have T6 [V250,V251,V252,V253,V254,V255,V256,V258,
7
turned on in order to limit the reverse voltage drop V260,V262,V264,V266]
uponT3.Dependently on the switching table used, the 8 [V303, V304, V305,V306, V307,V308]
maximum reverse voltage drop over the components 9 [V395, V396, V397, V398, V399, V400]
changes and a preliminary analysis must be done to [V488,V489,V490,V491,V492,V493,V494,V495,
10
choose the suitable component. Moreover, the switches V496,V498,V500,V502,V504,V506]
in the middle of the leg must carry twice the voltage of [V588,V589,V590,V591,V592,V593,V594,V595,
11
the others. V596,V598,V600,V602,V604,V606]
The DC bus voltage of 9-level inverter is split into nine [V676,V677,V678,V679,V680,V681,V682,V683,
levels by using eight DC capacitors C1,C2,C3,C4,C5,C6,C7 12
V684,V685,V686]
and C8 as shown in figure1. The voltage across the 13 [V686, V688, V690, V692, V694, V696]
switches is 1/8th of the DC bus voltage, switching losses
16
The space vector modulation diagram of 9-level inverter As shown on figure, DTC scheme needs only the output
is divided into six sextants, and each sextant is divided voltages and currents of the inverter which feeds the
into sixty-four triangles regions in order to show the induction motor, the instantaneous values of flux and
vectors nearest to the reference voltage as shown in torque in the machine are then calculated and the error
figure 2 and 3. can be gotten after compared with the referring torque
The voltage vectors of 9-level inverter are divided into and flux (Teref and  sref )
thirteen groups as shown in table II. The procedure of
selection of voltage vectors is similar to that of 3-level
inverter is discussed in [8].
Figure.4: Configuration of proposed DTC Scheme

The flux control is made by two-level hysteresis
comparator and C  defines the action wished on the
behaviour of the field. Also, a high level performance
torque is required and the torque control is controlled by
a hysteresis comparator built with eight lower bounds and
Figure.2: voltage vectors of 9-level MPC VSI eight upper know bounds and C  defines the action
wished on the behaviour of the torque [9]. The hysteresis
blocks are designed as it is shown in figure 5.
For the switching vector selection, it is necessary to know
the angular sector in which the actual flux is located. The
actual position of the stator flux can be determined by
equation (3), from the orthogonal flux components:
  s 
  tan 1   (3)
  s 
The flux position in the  and  Planes is quantified in
six sectors S of 600 degrees staring with the first sector
situated between -300 and 300.
Figure.3: Zoom of the area near to the reference

Space vector diagram of 9-level inverter in
Sextant-I
3. Representation of Proposed DTC Scheme

The basic requirements of DTC strategy is the estimation
of the stator flux and electromagnetic torque. The flux
evaluation can be carried out by different techniques
depending on the whether the rotor angular speed or
position which are measured or observed. For multilevel
applications, the voltage model is usually employed. The
stator flux can be evaluated by integrating from the stator
voltage equation
 s   U s  Rs i s t (1)
Further to calculation of the components of the flux, the
estimated torque is determined from the following
equation (2)
e   (i s  ds  is  qs ) (2)
Figure.5: Flux and Torque Hysteresis blocks
17
TABLE VI
Proposed Switching Table
S
CΦ CЃ 1 2 3 4 5 6 7 8 9 10 11 12
-8 684 694 685 696 680 686 681 688 682 690 683 692
-7 594 604 595 606 590 596 591 598 592 600 593 602
-6 494 504 495 506 490 496 491 498 492 500 493 502
-5 254 264 255 266 250 256 251 258 252 260 253 262
-4 184 194 185 196 180 186 181 188 182 190 183 192
-3 127 143 128 145 123 135 124 137 125 139 126 141
-2 307 399 308 400 303 395 304 396 315 397 306 398
-1 5 48 6 49 1 44 2 45 3 46 4 47
0 Zero Vector
+1 3 46 4 47 5 48 6 49 1 44 2 45
+2 305 397 306 398 307 399 308 400 303 395 304 396
+1 +3 125 139 126 141 127 143 128 145 123 135 124 137
+4 182 190 183 192 184 194 185 196 180 186 181 188
+5 252 260 253 262 254 264 255 266 250 256 251 258
+6 490 498 491 500 492 502 493 504 488 494 489 496
7 590 598 591 600 592 602 593 604 588 594 589 596
8 678 686 679 688 680 690 681 692 676 682 677 684
-8 696 680 686 681 688 682 690 683 692 684 694 685
-7 590 596 591 598 592 600 593 602 594 604 595 606
-6 496 491 498 492 500 493 502 494 504 495 506 490
-5 251 258 252 260 253 262 254 264 255 266 250 256
-4 188 182 190 183 192 184 194 185 196 180 186 181
-3 125 139 126 141 127 143 128 145 123 135 124 137
-2 397 306 398 307 399 308 400 303 395 304 396 315
-1 4 47 5 48 6 49 1 44 2 45 3 46
0 Zero Vector
+1 45 3 46 4 47 5 48 6 49 1 44 2
-1 +2 397 306 398 307 399 308 400 303 395 304 396 305
+3 139 126 141 127 143 128 145 123 135 124 137 125
+4 183 192 184 194 185 196 180 186 181 188 182 190
+5 262 254 264 255 266 250 256 251 258 252 260 253
+6 492 502 493 504 488 494 489 496 490 498 491 500
+7 602 593 604 588 594 589 596 590 598 591 600 592
+8 681 692 676 682 677 684 678 686 679 688 680 690
-8 692 684 694 685 696 680 686 681 688 682 690 683
-7 604 595 606 590 596 591 598 592 600 593 602 594
-6 506 490 496 491 498 492 500 493 502 494 504 495
-5 256 251 258 252 260 253 262 254 264 255 266 250
-4 188 182 190 183 192 184 194 185 196 180 186 181
-3 139 126 141 127 143 128 145 123 135 124 137 125
-2 398 307 399 308 400 303 395 304 396 315 397 306
-1 48 6 49 1 44 2 45 3 46 4 47 5
0 Zero Vector
+1 47 5 48 6 49 1 44 2 45 3 46 4
+2 308 400 303 395 304 396 305 397 306 398 307 399
0 +3 145 123 135 124 137 125 139 126 141 127 143 128
+4 186 181 188 182 190 183 192 184 194 185 196 180
+5 258 252 260 253 262 254 264 255 266 250 256 251
+6 498 491 500 492 502 493 504 488 494 489 496 490
+7 600 592 602 593 604 588 594 589 596 590 598 591
+8 690 681 692 676 682 677 684 678 686 679 688 680
18
Using the hysteresis comparator outputs flux C  and 1.5
torque CT and the stator flux sector S, the proper output

vector can be chosen to correct the error due to the 1
relation in the switching table. The switching
configuration is made step by step. The selection of a 0.5
Stator D-Q flux(Wb)

voltage vector at each cycle period Te is carried in order
to maintain the flux and torque within the limits of two 0
hysteresis bands. Several switching tables for three-level
inverter are presented in [10].
-0.5
A new switching table for the 9-level inverter selector
has been developed in Table III, to achieve an accurate
-1
control. In order to simplify, the mechanical rotor speed
will be considered when assigning the voltage vectors
needed at each one of those sectors. The speed of the -1.5
1 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09 1.1
stator flux linkage vector is given by the modulus of the Time(sec)
applied voltage vector. Thus, the voltage vectors will be Figure.7: D-Q axis flux response of 9-level Inverter fed
chosen according to the rotor speed [11]. Voltage vectors DTC IM drive.
with low amplitude will be chosen for lower speeds.
4. Simulation Results and Discussion
The proposed scheme is simulated in DTC with
15
multilevel control. The induction motor parameters are as
follows: Rs=4.85Ω, Rr=3.805Ω, Ls=274mH, Lr=274mH, 10
Lm=258mH, p=2, J=31g.m2, V=220V, power=1.5kW
and speed=1420rpm. All simulations have a sample time 5
Phase current(A)
for the control loop of 100μs; the voltage of the DC bus

is 514V. The requested space voltage vector, demanded 0
by the DTC strategy, is assured as shown by the line

voltage in Figure.13 Thus, if compared with a seven-level -5
DTC strategy [9]; the dv/dt applied to the motor

-10
terminals is greatly reduced. Furthermore, the
electromagnetic torque is also improved. It is possible to -15
reduce pulsation amplitude by approximately a factor of
2 when compared to a seven - level DTC strategy. 0.95 1 1.05
Time(sec)
1.1 1.15 1.2
To show the effectiveness of the DTC with 7-level and 9-

level inverters with SVPWM switching technique a Figure.8: Phase current response of 7-level Inverter fed
simulation work has been carried out on induction motor. DTC IM drive .
1.5
15
1
10
0.5
S ta to r D -Q flu x (A )
5
P has e Current(A )
0 0
-5
-0.5
-10
-1
-15
-1.5 0.95 1 1.05 1.1 1.15 1.2

1 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09 1.1
Time(Sec) Time(sec)
Figure.6: D-Q axis flux response of 7-level Inverter fed Figure.9: Phase current response of 9-level Inverter fed
DTC IM drive DTC IM drive
19
0.6
0.4
Electromagnetic Torque(N-M)
0.2
-0.2
-0.4
-0.6
0.4 0.405 0.41 0.415 0.42 0.425 0.43 0.435 0.44 0.445 0.45
Time(sec)
Figure.14:Stator line voltage harmonic spectrum of 9-
Figure.10: Torque error for 7-level inverter fed DTC IM level inverter
drive
0.1
0.05
Electromagnetic Torque (N-M)
-0.05
-0.1
-0.15
0.4 0.405 0.41 0.415 0.42 0.425 0.43 0.435 0.44 0.445 0.45
Time(sec) Figure.15: Stator line voltage harmonic spectrum of 9-
Figure.11: Torque error for 9-level inverter fed DTC IM level inverter
drive 1.04
600 1.03
1.02
400
Stator flux responce (wb)
1.01
200
Stator line voltage(v)
0
0.99
-200 0.98
0.97
-400 0.2 0.21 0.22 0.23 0.24 0.25 0.26 0.27 0.28 0.29 0.3
Time(sec)
-600
Figure.16:Stator flux error of 7-level Inverter fed
0.2 0.205 0.21 0.215 0.22 0.225
Time(sec)
0.23 0.235 0.24 0.245 0.25
DTC IM drive
Figure.12: Stator line voltage of 7-level inverter 1.01
1.008
600
1.006
400
Stator flux responce (wb)
1.004
200 1.002
Stator line voltage(v)
1
0
0.998
-200 0.996
0.994
-400
0.992
0.2 0.21 0.22 0.23 0.24 0.25 0.26 0.27 0.28 0.29 0.3
Time(sec)
-600
0.2 0.205 0.21 0.215 0.22 0.225 0.23 0.235 0.24 0.245 0.25
Time(sec) Figure.17: Stator flux error of 9-level Inverter fed
Figure.13: Stator line voltage of 9-level inverter DTC IM drive
20
In figure 6 and 7 D-Q axis flux response of the 7 and 9 - Nomenclature

level inverters are compared. It is seen that the  S  Stator flux linkage
performance of the 9-level inverter fed DTC IM drive has
lower ripple, so the proposed system is superior to US = Stator voltage expressed in the stator reference
control the flux with reduced ripple content. Figure 8 and frame
9 illustrate the phase current performance of 7-level and RS = Stator resistance
proposed inverter system, ,the phase current response 9- IS = Stator current
level inverter is fast recovery is obtained .Figure 10 and t = Standard deviation
11 shows the torque errors of the two systems, from the e = Instantaneous value of the electromagnetic torque
simulation results 7-level inverter torque error is 0.5 N-M
and 9-level inverter fed DTC IM drive torque error is
 = Derivation operator
reduced to 0.09 N-M, so from the proposed system is = Imaginary component of stator current
torque ripple minimization is possible. The stator line to
line voltages of 7-level and 9-level inverter system are is = Real component of stator current
illustrated in Figure 12 and 13.  ds = D-axis stator flux linkages
Figures 14, 15 shows the Stator line voltage harmonic
spectrum of 7 and 9-level inverter system. The 7-level  qs = Q- axis stator flux linkages
inverter THD is 7.81% and proposed system THD is Tref = Torque reference
reduced to the 5.99%. From the analysis 9-level inverter sref = stator flux reference
has reduced THD; it indicates that dv/dt applied to the
motor terminals is greatly reduced. Furthermore, the
electromagnetic torque is also improved. Figure 16 and 6. References
17 shows the stator flux errors response of the two [1] M. F. Escalante, J-C. Vannier and A. Arzandé.Flying
systems (7 and 9-level inverters). Capacitor Multilevel Inverters and DTC Motor Drive
From the above discussion, the proposed DTC IM drive Applications. IEEE Trans. On Industrial Electronics,
system behavior is good, even in extreme conditions like vol. 49, (No.4), pp. 809-815, 2002.
the reverse speed reference with nominal load torque [2] J. Rodriquez, J-S Lai and F. Z. Peng.Multilevel
applied. Reduction in ripple is observed in both Inverters: A survey of topologies, controls, and
electromagnetic torque and flux is due to the use of applications.IEEE Trans. On Industrial Electronics,
hysteresis controllers. vol. 49, (No.4), 2002.
[3] I. Takahashi and T. Nogushi. A new quick-response
and high efficiency control strategy of induction
5. Conclusions motor. IEEE Trans. On. IA, vol. 22, (No.5), pp. 820-
A nine–level inverter system for DTC induction motor 827, 1986.
drive is presented. DTC strategy of a nine–level Multi [4] G. Buja, D. Casadei and G. Serra. Direct torque
Point Clamped (MPC) voltage source inverter fed control of induction motor drives. Proc. IEEE
induction motor is implemented using Matlab/Simulink . International Symposium on Industrial Electronics,
Nine-level inverter has 729 switching states and there are vol. 1, pp.TU2-TU8, 1997.
106 effective vectors are possible. The proposed DTC IM [5]P. Vas, Sensor less Vector and Direct Torque Control.
drive scheme is capable for enough degrees of freedom to Oxford, U.K.: Oxford Univ. Press, 1998.
control both electromagnetic torque and stator flux with [6]D.Casadei, G.Grandi, G. Serra and A. Tani, Switching
very low ripple. strategies in direct torque control of induction
The following points are most attractive features of machines, ICEM 94, Vol. 2, pp. 204-209, 1994.
multilevel inverter fed DTC IM drive. [7] I. Messaïf, E.M. Berkouk, N. Saadia and A. Talha,
Application of Direct Torque Control Scheme for
 The output voltages with extremely low Induction Motor, International AMSE Conference
distortion and lower dv/dt . MS'05, 2005.
 They can operate with a lower switching [8]I.Messaïf, E.M. Berkouk, N. Saadia, Ripple Reduction
frequency. in DTC Drives by Using a Three-Level NPC VSI,
 As the number of levels increased 14th IEEE International Conference on Electronics,
the %THD in the motor phase voltage Circuits and Systems, Morocco, ICECS2007.
decreased,. [9]I. Messaïf, E.M. Berkouk, N. Saadia, An Improved
 The number of level increased the torque DTC strategy for induction machine control fed by a
ripple is reduced to minimum and the stator multilevel voltage source inverter, IEEE International
flux ripple is also minimized. Conference on Electronics, Circuits and Systems,
 For the proposed system has stator flux Morocco.
trajectory response is a circle and answer [10]B.P. McGrath,D.G. Holmes and T.A. Lipo,Optimised
the response is faster as compared to the 7- Space Vector Switching Sequences for Multilevel
level inverter. Inverters , IEEE Transactions on Power Electronics,
Vol.18 Issue:6,pp 1293 – 1301,2003.
[11]O.ChandraSekhar and K.ChandraSekhar, Space
vector modulation & fuzzy PID speed controller for
direct torque control induction motor drive. Journal of
21
Theoretical and Applied Information Technology.

Vol.35 no.2,pp 126 – 134, 2012.
[12]T.BrahmanandaReddy,K.Sri Gowri and Ch.Sai
Babu,GDPWM Algorithm for Direct Torque
Controlled Induction Motor Drive Using the Concept
of Imaginary Switching Times Automatic Control and
System Engineering ACSE",vol.8,no.3,pages37—
44,2008.
[13]N.R. Reddy,T.BrahmanandaReddy, J. Amarnath, and
D. Subba Rayudu ,Minimum Switching Loss PWM
Algorithms for Three-Phase Voltage Source Inverter
Fed Induction Motor Drives, Automatic Control and
System Engineering ACSE,vol.10,no.1,pages39-
48,2010.
Biographies
Mr.O.Chandrasekhar received his
B.Tech degree in Electrical &
Electronics Engineering from
JNTUH,India in 2005 and
M.Tech with power Electronics
and Electrical Drives from
Vignan’s Engineering
College,Vadlamudi, India in
2008.He has been with Vignan’s
Lara Institute of Technology and Science ,Vadlamudi as
Associate Professor. Presently he is a part-time research
student at J.N.T.U College of Engineering, Hyderabad-
500072, India, working towards his doctoral degree. His
Research interests are Power Electronics, Industrial
Drives.
Dr.K.ChandraSekhar received
his B.Tech degree in Electrical
& Electronics Engineering from
V.R.Siddartha Engineering
College, Vijayawada, India in
1991 and M.Tech with
Electrical Machines &
Industrial Drives from Regional
Engineering College, Warangal,
India in 1994. He Received the
PhD degree from the J.N.T.U College of Engineering,
Hyderabad, India in 2008. He is having 17 years of
teaching experience. He is currently Professor and Head
of Department, EEE, R.V.R & J.C.College of
engineering, Guntur, India. His Research interests are
Power Electronics, Industrial Drives & FACTS Devices.
22
Hybrid MPPT-Controlled LED Illumination Systems
Akram Abu-aisheh1 and Sameer Khader2

1
University of Hartford, West Harford, Connecticut, USA
2
Palestine Polytechnic University, Hebron, Palestine
Abstract Application is fast expanding with the requirements of

This paper presents a road map for the design and data storage and high efficiency requirements in supply
development of sustainable hybrid High Brightness Light distributions. Compact fluorescent lamps are becoming
Emitting Diode (HB LED) illumination systems with obsolete, and replaced with HB LED lamps driving the
maximum power point tracking (MPPT). The proposed efficiency of lighting systems. The system presented in
hybrid system design presents the foundation for future this paper expands on the continued growth of DC power
sustainable high efficiency illumination systems. The distribution in buildings, and driving this trend toward that
system design presented in this paper individually end LED-based lighting has been cited as a factor.
controls an array of HB LEDs. Each HB LED operates
in a different mode defined by the user. Photovoltaic 2. Hybrid LED Illumination
(PV) panels-based systems are used as the primary source The principal motivation for this hybrid system is the
of energy in this system while the electric grid is used as clear shift to DC systems with more use of alternative
the backup source that supplies power to the HB LEDs energy sources. The AC vs. DC battle raged when
only when the power of the PV system is not sufficient to Edison promoted DC power while George Westinghouse
supply all the system HB LEDs and the PV system felt that AC was the way to go. As we know, AC won the
batteries are depleted below the minimum level set in the battle, since it was so much easier to step the voltage up
system design. A matlab/ Simulink model is proposed for and down using transformers, and higher voltages greatly
studying the system behaviors and operation modes. reduce resistive loss. This paper explores the continued
growth of DC power, and the force driving the trend;
LED-based lighting has been cited as one major driver.
Keywords: MPPT, High Brightness LED; Hybrid;
Since photovoltaic panels are used to power the
Matlab/ Simulink; DC Chopper
illumination system, there is a need for a second source
of power for the system when the sun is out for a period
1. Introduction of time beyond the capability of the batteries to serve as a
The demands for utilizing alternative power sources have backup source. Some products that are appearing in the
increased due to rising oil prices and more stringent market use HB LED with build-in converter that can be
environmental regulations. Alternative energy and its connected directly to the AC line. These products take
applications have been heavily studied for the last decade, advantage of the high efficiency of HB LEDs, but they
and solar energy is the preferred choice in many do not use a sustainable source as the main supply for
applications [1]. Among solar energy applications, the those LEDs. The authors present this new hybrid HB
photovoltaic (PV) technology has received much LED-based illumination system design with automatic
attention, and it is being used in many applications [2&3]. transfer switch as given in figure 1. This hybrid system
This paper presents a sustainable hybrid high brightness uses solar as the primary source of energy, and it
LED Illumination System that can be used to replace switches to the AC line only when the primary source
current illumination systems in order to improve the can’t supply the required power to the system.
system efficiency and reliability. In the proposed system, In the system given in figure 1, LEDs are powered from
a Single-ended primary-inductor converter (SEPIC) DC- the solar panel during the daytime as the primary source,
DC converter is used to deliver solar energy via PV cell with the battery in charge mode, and they operate from
modules to a battery bank in charging mode during the the battery at night. If the battery is fully discharged, the
daytime. At night, it drives an LED lighting system. HB LED System operates from the AC line as the
The main reason for choosing HB LEDs as the secondary source. The design of DC-DC converters, DC-
illumination source in this system is due to their high AC rectifiers, and DC-AC inverters are well understood
efficiency as compared with incandescent and fluorescent and can be followed using senior level or graduate power
lamps. A second factor is that the world of DC electronics textbooks [4-8]
23
is operated by means of a controller circuit, i.e. an FPGA.

Figure 5 presents the basic control system layout used for
this DC illumination control strategy that gives the user
more flexibility and control than the flexibility and
control level available for standard AC powered
illumination used in incandescent or fluorescent lamps.
Figure 3: AC and DC Contactors Design to Control the LED

Figure 1: Hybrid High Brightness LED Illumination System Illumination System Switching
Since LEDs can handle voltage fluctuations, the system

can be simplified by using a full wave bridge rectifier
instead of the AC-DC converter. The simplified system
is given in figure 2 where components can be classified
into two different categories. The first one is high
tension parts which consist of photo cells and high power
LED, and the other is the low tension side which consists
of L, C, and battery. The operation of this buck-boost
converter has two operation modes: the first is the buck
mode for day time battery charging operation, and the
second is the boost mode for evening time lighting usage. Figure 4: Relay Design used to Control the Hybrid Illumination
Switching
Figure 5: FPGA-Controlled Illumination
Figure 2: Simplified Hybrid Illumination The FPGA is programmed using Hardware Description
Language (HDL). In this case VHDL is used. [VHDL is
This simplified system is sustainable for many a hardware description language used in electronic design
applications emerging due to the continued growth of DC automation to describe digital and mixed systems and
power distribution in buildings, and the principal force integrated circuits.] The pattern of the LEDs defined by
behind this trend is LED-based lighting which we have the user can be modified using this hardware language.
cited as one major driver. The main challenges in Basys Spartan 2 FPGA is used for this application due to
building such a system include analysis of a switching its flexibility, low cost, and ease of use.
circuit that controls alternating of power between the AC The application software used in our case is Adept which
and DC distribution systems, being fed from Solar panel is a powerful program that allows configuration and data
and switching between the two sources. Figure 3 transfer with Xilinx logic devices, and is used as an
presents one solution for this problem where an AC interface between Xilinx and the Spartan 2 FPGA. For
contactor and a DC contactor with mechanical interlock maximum intensity from the LEDs, the typical forward
are used, and figure 4 presents the relay control for the voltage of 3.9V, with forward current 700mA, is supplied
mechanically interlocked contactors. to the FPGA. A personal computer provided with Xilinx
The illumination system is designed to be used to software environment is used for programming the
individually control an array of HB LEDs. Each HB LED sequence of the LEDs defined by the user. The
operates in a different mode that is independently defined programming is done again using VHDL. The developed
by the user. The defined sequence of LED illumination software was tested using Digilent FPGA.
24
High brightness LEDs [9&10] can be driven at currents input voltage that varies from above to below the output
from hundreds of mA to more than an ampere, compared voltage. This type of topology is needed when the
with the tens of mA for other LEDs; however few of the voltage from an unregulated input power source such as
HB LEDs can produce over a thousand lumens. Since solar where the sun irradiation, temperature and weather
overheating is destructive, the HP LEDs may need to be changes directly affect the generated output voltage. The
mounted on a heat sink to allow for heat dissipation. If standard SEPIC topology [11&12] requires two inductors
the heat from an HB LED is not removed using a heat in additional to step-up transformer, making the power-
sink, the device will burn out in seconds. supply footprint quite large.
A single HB LED can often replace an incandescent bulb Photovoltaic (PV) cells are used to convert the sunlight
in a torch, or be set in an array to form a powerful HB into electrical energy. On the other hand, it is also an
LED system. LEDs can operate on AC power without the important issue to save the energy demand and increase
need for a DC converter. Each half-cycle part of the LED the energy efficiency [13-15]. High brightness light
emits light with the other half cycle being dark, and this emitting diodes (LEDs) [16&17] are becoming more
is reversed during the next half cycle. The efficacy of this widespread for the lighting applications such as
type of HB LED is typically around 40 LM/W. A large automobile safety and signal lights, traffic signals, street
number of LED elements in series may be able to operate lighting and so on. LED power circuitry is discussed
directly from line voltage. In 2009 Seoul Semiconductor thoroughly in the literature [18].
released a high DC voltage capable of being driven from In lighting applications with solar energy, the charger is
AC power with a simple controlling circuit. The low adapted to convert the solar irradiations for storing in the
power dissipation of these LEDs gives them more battery during the daytime. In the nighttime, a discharger
flexibility than the original AC LED design. In this is used to release energy in the battery and drive the LED
project the HB LEDs are powered from a DC source. lighting system. Low-power DC-DC converters can be
To control the LED from the FPGA, a driver circuit is used for the charger and discharger mode. Since the PV
needed to boost the power of the FPGA output. An LED voltage from the solar panel is unstable, the buck-boost
driver circuit is an electric circuit used to power a light- converter is more suitable for the charger circuit. This
emitting diode or LED. The circuit consists of a voltage converter can also be used in the discharger circuit.
source powering a current limiting resistor and an LED
connected in series. The HB LEDs used in our design 4. DC-DC Converter MPPT Control Design
have a constant current of 700mA and a supply voltage of and Analysis Using Computer Simulation
+3V. As the current has to be amplified to 700mA for Computer simulation is an important tool for future
each LED, two transistors are connected together so that illumination systems design. The HB LED-based
the current amplified by the first is amplified further by illumination system was simulated using a
the second transistor. The overall current gain is equal to matlab/Simulink environment [19]. In this section, we
the two individual gains multiplied together, i.e hFE = present the complete simulation model including the
hFE1 × hFE2, where hFE1 and hFE2 are the gains of the solar PV module, Maximum Power Tracking module
individual transistors. (MPPT), DC-DC boost converter and storage unit
(Battery ). These modules are going to be described as
3. Sustainable PV-Powered Illumination follows:
Since photovoltaic panels are used to power the HB LED
illumination systems, there is a need for the development
of the DC converter system to power the LEDs. To 4.1 Characteristics of PV Array
satisfy this requirement, a forward converter with PV- Basically, PV cell is a P-N semiconductor junction that
based LED applied in lighting systems was used. In the directly converts light energy into electricity. It has the
proposed DC supply system, the SEPIC is used to deliver equivalent circuit shown in Figure1 as represented in [20-
solar energy via PV cell modules to a battery bank in 22].
charging mode during the daytime. During the nighttime,
the converter drives an LED lighting system.
Figure 7 illustrates the principle PV-SEPIC- LED circuit
applied in street illumination, where the produced PV
voltage is stored in a battery bank throughout the
charging unit during the daytime, at night the discharger Figure 6: Equivalent circuit for PV cell
activates and the LEDs are energized with appropriate
voltage throughout the step-up transformer and full Where Iph represents the cell photocurrent; Rp and Rs are
bridge rectifier. the intrinsic shunt and series resistance of the cell
SEPIC circuits find widespread application when the respectively; Id is the diode saturation current; Vo and Io
input voltage fluctuates above and below average value,
are the cell output voltage and current respectively.
while the output voltage must be kept at a constant value
with minimum tolerances. One of most important The following are the simplified equations describing the
applications of SEPIC circuit is the integration with the cell output voltage and current:
photovoltaic system (PV system) and illumination load of A.K.T c  Iph  Id  Io 
series and parallel connected LEDs. Vo  ln    Rs.Io (1)
The SEPIC converter is a DC/DC converter topology that q  Io 
provides a positive regulated output voltage from an
25
 q . Vo / Ns  5
I-V performance
Io  N p ( Iph  Id  e A . K .Tc  1  (2) 4.5 1200W/m2
  4
1000W/m2
3.5
3 q . Eg  1 1 
 Tc 
3
   800W/m2
I d  I or   .e
B . K  Tr Tc 
(3)
Ipv,A
2.5
 Tr  2
600W/m2
1.5
I ph  N p .I sc . n  I t ( T c  T r ) 
400W/m2
(4) 1
0.5
Where, K- Boltzmann constant; Np and Ns are the 0

0 5 10 15 20 25
number of parallel and series connected cells respectively; Vpv, V
b) I-V Performance of PV module.

Eg is the band gap of the semiconductor; Tc and Tr are the
cell and the reference temperature respectively in Kelvin, 90
Power performance
A and B are the diode ideality factors with values varies 80

1200W/m2
between 1 and 2; n is the normalized insulation; Isc is 70
the short circuit current given at standard condition; It 60
and Ior are constants given at standard conditions. 50
Ppv, W
40
4.1 Photovoltaic I-V Performance 30
In order to study the I-V performance of the PV circuit 20
and to look for appropriate dc chopper for boosting up 10

400W/m2
the output voltage to predetermined value it is necessary 0

0 5 10 15 20 25
to illustrate the obtained PV voltage and current for Vpv,V
boost chopper according to specifications given in table 1 c) Power performance of proposed module
at reference irradiation 1000W/m2. Figure 7: PV Array module & main charaterestics.
The PV Array voltage can be obtained by multiplying the
module voltage and current by Nsm and Npm that Figure 7 illustrates the proposed PV array with R load
represents number of series and parallel connected presenting the equivalent resistance of complete LED set,
modules respectively. where the obtained results for different variation levels
are presented. From these performances it is shown that
Table 1: Data specification for PV Array. the total output PV voltage and current varies according
to irradiation level with approximated 65W maximum
q K Iph Id RS RP TC
power at G=1000W/m2.
1.602e- 1.38e-
4 A 0.2mA 23m 0.65k 25C
19 C 23J/K
NS NP VO VOC ISC VMPP IMPP 4.2 Maximum power point derivation
38 4 0.6V 21.5 V 4.7A 17.5V 3.7A In order to operate the module at maximum extracted
NSm NPm Rload power, it is necessary to calculate the coordinates of the
1 1 4.5( Set of Series connected LED maximum power point (VMPP, IMPP). For this, and to
Diodes) simplify the model in Simulink, the coordinates of the
maximum power point are given by the following
equations:
C ontinuous
powergui
G_T
IMPP  Iph  Id. e Vmpp / VT  1  (5) 
 Iph  Im pp VOC / VT
1
T
5
T_var
T +
- v
VMPP  VT. ln1 
Iph
e

1  (6)  
1 1200 G
V2
 
G G_var
i
Lf Output Where VT and Voc are the thermal voltage
6 Ns +Vpv + - and the cell open circuit voltage respectively
Ns I
1 Np
and given by:
Np + K .T c
Cf R-LED -
v
VT 
GND
V1 q (7)
PV Array
 I sc 
V OC  V T . ln   1
 I d 
The maximum power that can be obtained can be
a) Matlab/Simulink module for PV Array
expressed as follows:
P MPP  V MPP . I MPP (9)
26
Obtaining ng the maximum obtained power form the

PV for given solar irradiation requires controlling the duty
cycle D of DC Buck-Boost Chopper toward operating the
chopper in mode where the extracted from the system
power will be maximum. There are several approaches
used for realizing MPPT operation mode, one of the most
popular and simple approach de is called Perturbation &
Observation model (P&O MPPT model).
The mathematical equations realized MPPT approach are
expressed as follows:
1
Vo  Vpv .
1 D
Po  Vo / RI
2
dPo  Po ( k  1)  Po ( k );
dVo  Vo ( k  1)  Vo ( k ); (10)
(dPo/dVo)  ?.....  D  ?
  0  D  D  D
  0  dVo 
  0  D  D  D
dPo 
 0  dVo   0  D  D  D
 
  0  D  D  D
While figure8 illustrates the Power performance and the
flowchart procedure for obtaining the appropriate duty b) Flowchart for tracking procedure
cycle needed to operate the chopper at maximum obtained
Figure 8: P&O MPPT tracking method.
power.
a) Boost DC Chopper
a) Maximum Power Point Tracking procedure

b) Mode#1: IGBT=ON
The proposed mathematical equations and corresponds
flowchart can be summarized with operation logic as
follow:
 (dPo/dVo)<0  operation at the right hand side of
the curve (fig.8a), for D Dmin , then V must be
decreased ( moving toward decrease).
c) Mode#2: IGBT=OFF
 (dPo/dVo>0)  operation at the left hand side of
the same curve , for D Dmax1). Then V must be
increased (moving toward increase).
4.3: Boost DC Chopper

The output PV voltage in most of applications is requred
to be boost up by applying Boost DC chopper [6]as d) Load voltage & current at two operation modes
shown in figure 9, where the principle electrical circuit
and the operation mode at continuous current mode Figure 9: Equivalent circuit for boost converter operating in CCM
(CCM) are illustrated.
27
4.4: Matlab Simulation of Proposed Model 2400

Irradiation vs simulation time
The complete electrical circuit for LED illumination 2200

system is simulated in mat lab/simulink environment and 2000
illustrated in 10
1800
Irradiation, W/m2
PV_Current
1600
MPPT_Duty _Cy cle
PV_Voltage
PV_Specifications.m 1400
MPPT
PV_perf
Discrete, Dbl_clk_for_init
Ts = Te s. 1200
pow ergui
MPPT Duty Cycle
Pv _Voltage 1000
Duty
Ppv
800
Pv _Current
PV_Current
600
Rate1 Ipv
V+PV
298 RT Ipv
T RCable 400
D1 0 2 4 6 8 10 12
Vpv +Ve_Load
T,[K]
+ R-LED Vout
V+pv Simulation interval x 10
4
+
+
Rate2 G m Ir v
-
V-pv _
Pout
s
-
+
RT V-PV I2
a) Solar irradiations at various daytime intervals.

PV
-Ve_load
-
Battery
i
Radiation
Vpvin-out
buck-Boost_Converter I_load 10
Duty cCycle vs simulation time
a) Load voltage & current at two operation modes 1
D1
0.9
L 0.8
3
V+PV 1 0.7
D2 +Ve_Load
1 PWM_IN PWM_OUT
[Vpv]
0.6
Duty Cycle
[Vpv] Duty
PG From
D
g
Goto
Q_ch
0.5
R1
C_Output1
+ v
S
+
- v C_Output
- 0.4
VM Vout
V M1 +
Pv_Voltage
v 2 1
- 0.3
Vpv1 Gain_Divider
R2
0.2
-Ve_load
4 2 0.1
V-PV
RSh
0
Power2 Pv_Current 0 2 4 6 8 10 12
+
Vpv
v
-K- 2
Simulation interval x 10
4
Gain_Voltage_to_current
b) Chopper duty cycle.

b) DC Chopper simulation module
PV Array current vs simulation interval
8
PV_Current
1 7
Current_PV PV_Power
sat_Current MPPT_Duty_Cycle 6
ZOH 1/z
MATLAB
1
Function 5
PV_Voltage
Rate1 MPPT_M_Code Rate
2 4
Ipv, A
sat_voltage 3
1
Scope
0
c) MPPT module -1
Fig.10: Complete Simulation model of PV- LED illumination system. -2

0 2 4 6 8 10 12
Simulation interval x 10
4
4.5: Simulation Results c) Array current of The PV pannel

There are various PV and chopper paramtetrs that can be
120
described; some of these are illustrated in figure 11 as Pmpp @G=2000
follows: 100
 Solar irradiation: Figure 11(a) illustrates the time 80
profile of solar irradiation levels for various periods 60

Array parameters
of daytime in order to show the behaviors of the Pmpp @G=600

circuit according to these levels. 40
Ppv Vpv
 Chopper duty cycle: Figure 11(b) illustrates the time 20

profile of calculated dudty cycle , where its shown
that for each irradiation level the duyt cycle increases 0
which cause further increase until maximum obtaied -20

Ipv
power is achieved.
-40
 PV array current: Figure 11(c) illustrates the time 0 1 2 3 4 5 6 7
Simulation interval 4
profile of PV current, where its shown that high x 10
irradiation produces large current and power . d) PV voltage, current and power
28
Array & output voltage [7] Mohan, Undeland, and Robbins, "Power Electronics
70
Converters Applications and Design", John Wiley
60 and Sons, Inc. 2003.
Vout
[8] Shaffer, “Fundamentals of Power Electronics with
50
Matlab”, ISBN# 9781584508526
40
[9] Kingbright
Vpv & Vout, V
http://www.kingbrightusa.com/default.asp
30
Vpv
[10] Lumex http://www.lumex.com/marketingdownloads
[11] Jeff Fallin, "Designing DC/DC converters based on
20
SEPIC topology", Analogue Applications Journal,
10 Power Management, Texas Instruments Incorporated,
4Q, 2008, pp. 18-23
0
[12] Lin B.R. & Huang C.L., "Analysis and
-10
implementation of an integrated SEPIC -forward
0 2 4 6
Simulation interval
8 10 12
4
converter for photovoltaic-based light emitting diode
x 10
lighting", IET Power Electronics, 2009, Vol.2, Iss.6,
e) PV and chopper voltage pp.635-645.
Fig.11: Circuit performances at various irradiations [13] Hodge B. K., "Alternative Energy Systems &
applications", John Willey & Sons, Inc., 2010,
pp.114-148.
 PV voltage, current and power: Figure 11(d) [14] Chuang Y.C. & Ke Y.L.," High efficiency battery
illustrates the time profile of PV voltage , current and charger with a Buck zero-current switching pulse
power , where its shown how the power being width modulated converter", 2008, Vol.1, Iss.4,
affected by increasing the irradiation level. pp433-444.
[15] Jovcic D., "Step-up DC-DC converter for megawatt
 Input-output voltage: Figure 11(e) illustrates the size applications", IET Power Electronics, 2009,
time profile of PV input voltage and output chopper Vol.2, Iss.6, pp.675-685.
voltage, where its shown how the input voltage varies,
[16] Redley R., " Analyzying the Sepic converters",
while the output voltage is boosted up and fixed at
Power System Design Europe, November, 2006,
appropriate for the elimination system value.
pp.14-18.
5. Conclusion [17] Bisogno F.E., Nittayarumphong S., Radecker M.,
A Sustainable hybrid MPPT-Controlled HB LED-Based Doprado R.N., " A line power-supply for LED
Illumination System was developed. This illumination lighting using piezoelectric transformers in class-E
system can be used for many lighting applications since it topology", proc. IEEE IPEMC'06, 2006, Vol.2, pp1-
is more efficient and reliable than existing traditional 5.
lighting systems based on incandescent or fluorescent [18] Joseph T. Verdeyen. “Laser Electronics” 3rd edition
lamps. While the use of HB LEDs adds to the initial cost Pearson Prentice Hall, 1995.
of the system, it will be paid off in the long term due to [19] Matlab and Simulink, The Mathworks, Inc., version
their higher reliability, flexible control, and long life time. R2010, http://www.mathworks.com
A simulink model for this illumination system was built [20] Atlas H., Sharaf A.M.," A Photovoltaic array
in order to study the system behaviors at various simulation model for Matlab-simulink GUI
irradiation level, the effect of MPPT system and environment, IEEE, Trans., 2007, pp.341-345.
generated duty cycle on chopper operation, which in turn [21] Chouder A., Silvester S., Malek A., " Simulation of
causes better utilization, effecienecy enehancement and photovoltaic grid connected inverter in case of grid-
loss reduction. failure", Revue des energes Renouveables Vol. 9,
No4, 2006, pp.285-296.
6. References [22] Buresch M.," Photovoltaic energy systems design
[1] Donald L. Steeby, “Alternative Energy Sources and and Installation", McGraw-Hill, New York, 1983.
Systems,” Delmar Cengage Learning, 2012.
[2] Thomas Kessell, “Introduction to Solar Principles,”
Pearson Education, Inc. 2011.
[3] Peter Gevorkian, "Large-Scale Solar Power System
Design", Mc Graw Hill Higher Education, 2011.
[4] Issa Batarseh, "Power Electronic Circuits", John
Wiley and Sons, Inc. 2004.
[5] Mohammad H. Rashid, "Power Electronics Circuits
Devices and Applications", Pearson Education, Inc.
2004.
[6] Daniel W. Hart, "Power Electronics", Mc Graw Hill
Higher Education, 2010.
29
Biographies
AKRAM ABU-AISHEH is an
Associate Professor of Electrical
and Computer Engineering at the
University of Hartford where he
has served as the assistant chair of
the Electrical and Computer
Engineering Department and
director of the electronic and
computer engineering technology
program for two years. Dr. Abu-aisheh has a doctorate in
Optical Communications from the Florida Institute of
Technology and Master of Science and Bachelor of
Science degrees in Electrical Engineering from the
University of Florida. His research interests include
Fiber Optic Communications, Solar Energy, Power
Electronics, and Engineering Education. He has
published a book, a book chapter, and several
international journals and conference papers. Dr. Abu-
aisheh may be contacted at abuaisheh@hartford.edu
SAMEER KHADER is an
Associate Professor of Electrical
and Computer Engineering at
Palestine Polytechnic University
(PPU) – Palestine. During the
2010-2011 year he was a visiting
professor at the University of
Hartford, CT, USA spending his
sabbatical year. He is a director
of Power Electronics & Signal Processing Research Unit
at PPU. Before that he served for ten years as university
academic provost, dean of college and Chair of the
Electrical & Computer Engineering Department at PPU.
His research interests include Electrical machines,
Electrical drives, Power Electronics Converters,
Renewable Energy Sources and Smart Grids, in addition
to engineering education. He has several publications in
international journals and conferences. Dr. Khader may
be contacted at sameer@ppu.edu
30
A sophisticated Space Vector Pulse Width Modulation Signal Generation for

Nine-Level Inverter system for Dual-Fed Induction Motor Drive
G. Sambasiva Rao, K. Chandra Sekhar

Dept. of Electrical Engg, R.V.R & J.C. College of Engineering, Guntur, India
sambasiva.gudapati@gmail.com
Abstract T*as, T*bs, T*cs The time equivalents of the voltage

A sophisticated space vector modulation scheme for a magnitudes
nine-level inverter system for dual-fed induction motor Tfirst_cross,
drive, using only the instantaneous sampled reference Tsecond_cross,
signals is presented in this paper. The dual-fed structure Tthird_cross The time intervals, at which the
is realized by opening the neutral-point of the sinusoidal reference signals cross
conventional squirrel cage induction motor. The nine- the triangular carrier signals for the
level inversion is obtained by feeding the dual-fed first time, second time and third
induction motor with symmetrical four-level inverter time
from one end and symmetrical three-level inverter from T0 The time interval of the start and
other end. The proposed space vector pulse width end space vector
modulation technique does not require the sector Tmiddle The time interval for the middle
information and look-up tables to select the appropriate inverter space vectors
switching vectors. The inverter leg switching times are
directly obtained from the instantaneous sampled 1. Introduction
reference signal amplitudes and centers the switching The two most widely used pulse width modulation
times for the middle space vectors in a sampling time (PWM) schemes for multilevel inverters are the carrier-
interval, as in the case of conventional space vector pulse based sine-triangle PWM (SPWM) scheme and the space
width modulation. vector PWM (SVPWM) scheme. The SPWM schemes are
more flexible and simpler to implement, but the maximum
Keywords: dual-fed induction motor, middle space peak of the fundamental component in the output voltage
vectors, nine-level inverter, sampled sinusoidal reference is limited to 50% of the DC link voltage [1, 2]. The
signals, space vector PWM. maximum peak of the fundamental component in the
output voltage obtained with space vector modulation is
Nomenclature 15% greater than with the sine-triangle modulation
Edc DC link voltage of an equivalent scheme [2-5]. But the conventional SVPWM scheme
conventional single two-level requires sector identification and look-up tables to
Inverter determine the timings for various switching vectors of the
Ts Sampling time inverter, in all the sectors [3, 4]. This makes the
EA3A5, EB3B5, EC3C5 Motor phase voltages implementation of the SVPWM scheme quite
Emax, Emin The maximum and minimum complicated. It has been shown that, for two-level
magnitudes of the three sampled inverters, a SVPWM like performance can be obtained
sinusoidal reference signals with a SPWM scheme by adding a common mode voltage
Eoffset1, Eoffset2 common offset voltages of suitable magnitude, to the sinusoidal reference signals
E*AN, E*BN, E*CN The modified reference signals [4 - 6].
after addition of Eoffset1 The SPWM scheme, when applied to multilevel inverters,
t1, t2 The time durations for the middle uses a number of level-shifted carrier signals to compare
inverter space vectors with the sinusoidal reference signals [7]. The SVPWM for
t01, t02 The time durations for the start and multilevel inverters [8 - 11] involves mapping of the outer
end inverter space vectors sectors to an inner sub-hexagon sector, will be very
Tx-cross(x=a, b, c) The time interval, at which the x- complex, as a large number of sectors and inverter vectors
phase reference signal, E*XN are involved.
crosses the triangular carrier signal
31
A modulation scheme is presented in [12], where a fixed and INV-3 in cascade. The Inverter-B is composed of
common mode voltage is added to the reference signal two conventional two-level inverters INV-4 and INV-5 in
throughout the modulation range. It has been shown [13] cascade. The DC link voltages of INV-1, INV-2, INV-3,
that this common mode voltage addition will not result in INV-4 and INV-5 are (2/8)Edc, (2/8)Edc, (2/8)Edc,
a SVPWM-like performance, as it will not centre the (1/8)Edc and (1/8)Edc respectively, where Edc is the DC
middle space vectors in a sampling interval. The common link voltage of an equivalent conventional single two-
mode voltage to be added in the reference phase voltages, level inverter drive.
to achieve SVPWM-like performance, is a function of the The leg voltage EA3n of phase-A attains a voltage of
modulation index for multilevel inverters [13]. A (2/8)Edc if (i)The top switch S31 of INV-3 is turned on
SVPWM scheme based on the above principle has been (Figure 1) and (ii) The bottom switch S24 of INV-2 is
presented [14], where the switching time for the inverter turned on. The leg voltage EA3n of phase-A attains a
legs is directly determined from sampled sinusoidal voltage of (4/8)Edc if (i) the top switch S31 of INV-3 is
reference signal amplitudes. But it involves region turned on (ii) The top switch S21 of INV-2 is turned on
identifications based on modulation indices. While this and (iii)The bottom switch S14 of INV-1 is turned on. The
SVPWM scheme works well for a three-level PWM leg voltage EA3n of phase-A attains a voltage of (6/8)Edc
generation, it cannot be extended to multilevel inverters of if (i) the top switch S31 of INV-3 is turned on (ii) The top
levels higher than three, as the region identification switch S21 of INV-2 is turned on and (iii)The top switch
becomes more complicated. A carrier-based PWM S11 of INV-1 is turned on. The leg voltage EA3n of phase-
scheme has been presented [15], where sinusoidal A attains a voltage of zero volts if the bottom switch S34
references are added with a proper offset voltage before of the INV-3 is turned on. Thus the leg voltage EA3n
being compared with carriers, to achieve the performance attains four voltages of 0, (2/8)Edc, (4/8)Edc and
of a SVPWM. The offset voltage computation is based on (6/8)Edc, which is basic characteristic of a 4-level
a modulus function depending on the DC link voltage, inverter. Similarly the leg voltages EB3n and EC3n of
number of levels and the sinusoidal reference signal phase-B and phase-C attain the four voltages of 0,
amplitudes. A SVPWM scheme is presented [18], where (2/8)Edc, (4/8)Edc and (6/8) Edc.
the switching time for the inverter legs is directly The leg voltage EA5n’ of phase-A attains a voltage of
determined from sampled sinusoidal reference signal (1/8)Edc if (i)The top switch S51 of INV-5 is turned on
amplitudes for five-level inverter where two three-level and (ii) The bottom switch S44 of INV-4 is turned on. The
inverters feed the dual-fed induction motor. A carrier leg voltage EA5n’ of phase-A attains a voltage of (2/8)Edc
based SPWM scheme is presented [19, 20] for five-level if (i) The top switch S51 of INV-5 is turned on and (ii)
and nine-level inverters. Qualitative SVPWM schemes are The top switch S41 of INV-4 is turned on. The leg voltage
presented for five-level inverters in [21, 22]. EA5n’ of phase-A attains a voltage of zero volts if the
The objective of this paper is to present an bottom switch S54 of the INV-5 is turned on. Thus the leg
implementation scheme for PWM signal generation for voltage EA5n’ attains three voltages of 0, (1/8)Edc and
nine-level inverter system for dual-fed induction motor, (2/8)Edc, which is basic characteristic of a 3-level
similar to the SVPWM scheme. In the proposed scheme, inverter. Similarly the leg voltages EB5n’ and EC5n’ of
the dual-fed induction motor is fed with symmetrical phase-B and phase-C attain the three voltages of 0,
four-level inverter from one end and symmetrical three- (1/8)Edc and (2/8)Edc.
level inverter from other end. The PWM switching times Thus, one end of dual-fed induction motor may be
for the inverter legs are directly derived from the sampled connected to a DC link voltage of either zero or (2/8)Edc
amplitudes of the sinusoidal reference signals. A simple or (4/8)Edc or (6/8)Edc and other end may be connected
way of adding an offset voltage to the sinusoidal to a DC link voltage of either zero or (1/8)Edc or
reference signals, to generate the SVPWM pattern, from (2/8)Edc. When both the inverters, Inverter-A and
only the sampled amplitudes of sinusoidal reference Inverter-B drive the induction motor from both ends, nine
signals, is explained. The proposed SVPWM signal different levels are attained by each phase of the
generation does not involve checks for region induction motor. The nine levels generated for phase-A
identification, as in the SVPWM scheme presented in are shown in Table 1.
[14]. Also, the algorithm does not require either sector
identification or look-up tables for switching vector TABLE 1
THE NINE LEVELS REALIZED IN THE PHASE-A WINDING
determination as are required in the conventional
multilevel SVPWM schemes [10, 11]. Thus the scheme is Leg- Leg- Motor phase voltage Level
voltage voltage EA3A5 = EA3n - EA5n'
computationally efficient when compared to conventional
of phase A, of phase A,
multilevel SVPWM schemes, making it superior for real- EA3n EA5n'
time implementation. 0 (2/8) Edc -(2/8) Edc Level 1
0 (1/8)Edc -(1/8)Edc Level 2
2. Nine-level inverter scheme for the dual- 0 0 0 Level 3
fed induction motor (2/8) Edc (1/8) Edc (1/8)Edc Level 4
(2/8) Edc 0 (2/8)Edc Level 5
The power circuit of the proposed drive is shown in
(4/8) Edc (1/8) Edc (3/8)Edc Level 6
Figure 1. A symmetrical four-level inverter, Inverter-A (4/8) Edc 0 (4/8) Edc Level 7
and a symmetrical three-level inverter, Inverter-B feed (6/8) Edc (1/8) Edc (5/8) Edc Level 8
the dual-fed induction motor. The inverter-A is composed (6/8) Edc 0 (6/8) Edc Level 9
of three conventional two-level inverters INV-1, INV-2
32
Figure 1. Schematic circuit diagram of the proposed 9-level inverter drive scheme.
3. Proposed SVPWM in linear modulation voltage Eoffset1, are shown in figure 2 along with eight
range triangular carrier signals T1 to T8. Each time a sinusoidal
For two-level inverters, in the SPWM scheme, each reference signal crosses the triangular carrier signal, it
sinusoidal reference signal is compared with the triangular causes a change in the inverter switching state. The
carrier signal and the individual phase voltages are changes in phase voltage and their time intervals are
generated [1]. To attain the maximum possible peak shown in Figure 3 in a sampling time interval Ts. The
amplitude of the fundamental phase voltage, a common sampling time interval Ts can be split into four time
offset voltage, Eoffset1 is added to the sinusoidal intervals t01, t1, t2 and t02. The time intervals t01 and t02 are
reference signals [5, 12], where the magnitude of Eoffset1 the time durations for the start and end inverter space
is given by vectors respectively and the time intervals t1 and t2 are the
time durations for the middle inverter space vectors
(active space vectors), in a sampling time interval Ts. It
Eoffset1  ( E max  E min) / 2 (1)
should be observed from Figure 3 that the middle space
vectors are not centered in a sampling time interval Ts.
Where Emax and Emin are the maximum and minimum Because of the level-shifted eight triangular carrier signals
magnitudes of the three sampled sinusoidal reference (Figure 2), the first crossing and the last crossing of the
signals respectively, in a sampling time interval. The sinusoidal reference signal cannot always be the minimum
addition of Eoffset1, results in the active space vectors and the maximum magnitude of the three sampled
being centered, making the SPWM scheme equivalent to sinusoidal reference signals, in a sampling time interval.
the SVPWM scheme [3]. In a sampling time interval, the Thus the offset voltage, Eoffset1 is not sufficient to center
sinusoidal reference signal which has lowest magnitude the middle inverter space vectors, in a multilevel PWM
crosses the triangular carrier signal first, and causes the system (Figure 3). Hence an additional offset (offset2) has
first transition in the inverter switching state and the to be added to the sinusoidal reference signals of Figure 2,
sinusoidal reference signal, which has the maximum so that the middle inverter space vectors can be centered
magnitude, crosses the triangular carrier signal last and in a sampling time interval, same as a two-level SVPWM
causes the last switching transition in the inverter system [3].
switching states in a two-level SVPWM scheme [5, 13].
For nine-level inverter, the modified sinusoidal reference
signals (E*AN, E*BN and E*CN) after addition of offset
33
4. Determination of the offset voltage for a

nine-level inverter The time intervals, Tfirst_cross, Tsecond_cross and
Figure 2 shows modified sinusoidal reference signals and Tthird_cross, directly decide the switching times for the
eight triangular carrier signals used for PWM generation different inverter voltage vectors, forming a triangular
for nine-level inverter. The modified sinusoidal reference sector, during one sampling time interval Ts. The time
signals are given by intervals for the start and end space vectors, are t01=
Tfirst_cross, t02= (Ts − Tthird_cross), respectively
(Figure 3). The middle space vectors are
E * AN  E AN  Eoffset1
centered by adding a time offset, Toffset2 to Tfirst_cross,
E * BN  E BN  Eoffset1 (2) Tsecond_cross and Tthird_cross. The time offset, Toffset2
E *CN  E CN  Eoffset1 is determined as follows. The time interval for the middle
where EAN, EBN and ECN are the sampled amplitudes of inverter space vectors, Tmiddle, is given by:
sinusoidal reference signals during the current sampling
time interval and Eoffset1 is calculated from equation (1). Tmiddle  Tthird _ cross  Tfirst _ cross (6)
The time interval, at which the A-phase sinusoidal
reference signal, E*AN crosses the triangular carrier signal, The time interval of the start and end space vector is
is termed as Ta-cross (Figure 4). Similarly, the time
intervals, when the B-phase and C-phase sinusoidal T0  Ts  Tmiddle (7)
reference signals, E*BN and E*CN cross the triangular
carrier signals, are termed as Tb-cross and Tc-cross Thus the time interval of the start space vector is given by
respectively. Figure 4 shows a sampling time interval
when the A-phase sinusoidal reference signal is in the T0 / 2  Tfirst _ cross  Toffset 2
triangular carrier region T7 while the B-phase sinusoidal
Therefore
reference signal and C-phase sinusoidal reference signal
are in carrier region T8 and T1 respectively. As shown in
Figure 4, the time interval, Ta-cross, at which the A- Toffset 2  T0 / 2  Tfirst _ cross (8)
phase sinusoidal reference signal crosses the triangular
carrier signal is directly proportional to the phase voltage In this way, we can obtain offset voltages to be added for
amplitude, (E*AN − 3Edc ⁄8). The time interval, Tb-cross, remaining samples during the time period of sinusoidal
at which the B-phase sinusoidal reference signal crosses reference signal and for different modulation indices.
the triangular carrier signal, is proportional to (E*BN +
4Edc ⁄8) and the time interval, Tc-cross, at which the C-
phase sinusoidal reference signal crosses the triangular
carrier signal, is proportional to (E*CN ). Therefore
 Ts 
Ta  cross  ( E * AN 3Edc / 8)   T * as  3Ts
 Edc / 8 
 Ts 
Tb  cross  ( E *BN 4 Edc / 8)   T * bs  4Ts
 Edc / 8 
 Ts  (3)
Tc  cross  ( E *CN )   T * cs
 Edc / 8 
Where T*as, T*bs and T*cs are the time equivalents of the
voltage magnitudes. The proportionality between the time
equivalents and corresponding voltage magnitudes is
defined as follows [6]:
( Edc / 8) / Ts  E * AN / T * as
( Edc / 8) / Ts  E * BN / T * bs
( Edc / 8) / Ts  E *CN / T * cs
( Edc / 8) / Ts  Eoffset1 / Toffset1 (4)
The time interval, at which the sinusoidal reference

signals cross the triangular carrier signals for the first Figure 2. Modified sinusoidal reference signals and triangular carrier
time, is termed as Tfirst_cross. Similarly, the time signals for a nine-level PWM scheme
intervals, at which the sinusoidal reference signals cross
the triangular carrier signals for the second and third time,
are termed as, Tsecond_cross and Tthird_cross
respectively, in a sampling time interval Ts.
Tfirst _ cross  min(Ta  cross, Tb  cross, Tc  cross)

T second _ cross  mid(Ta  cross, Tb  cross, Tc  cross)
Tthird _ cross  max(Ta  cross,Tb  cross,Tc  cross) (5)
34
simulation studies. Figure 5(a) shows the total offset

voltage to be added to sinusoidal reference signals to
make SPWM equivalent to the SVPWM in the lowest
speed range which corresponds to two-level mode when
modulation index is 0.1, figure 5(b) shows the A-phase
sinusoidal reference signal after offset voltage is added
and Figure 5(c) shows the motor phase voltage (EA3A5).
Figure 6 to Figure 12 show the motor waveforms in the
next speed ranges which corresponds to three-level mode
to nine-level mode when modulation indices are 0.2, 0.3,
0.42, 0.53, 0.62, 0.74 and 0.85 respectively. It can be
observed that the motor phase voltage during 9-level
operation is very smooth and close to the sinusoid with
lower harmonics.
6. Conclusion
A modulation scheme of SVPWM for nine-level inverter
system for dual-fed induction motor drive is presented.
The centering of the middle inverter space vectors of the
SVPWM is accomplished by the addition of an offset
voltage signal to the sinusoidal reference signals, derived
from the sampled amplitudes of the sinusoidal reference
signals. The SVPWM technique, presented in this paper
does not require any sector identification, as is required in
conventional SVPWM schemes. The proposed scheme
Figure 3. Inverter switching vectors and their switching time durations eliminates the use of look-up table approach to switch the
during sampling time interval Ts appropriate space vector combination as in conventional
SVPWM schemes. This reduces the computation time
required to determine the switching times for inverter
legs, making the algorithm suitable for real-time
implementation.
Figure 4. Determination of the Ta-cross, Tb-cross and Tc-cross

during sampling interval Ts
5. Simulation results and discussion

The proposed SVPWM scheme is simulated using
MATLAB environment with open loop E/f control for
different modulation indices. The speed reference is
translated to the frequency and voltage commands
maintaining E/f. The modified three reference sinusoidal
Figure 5. (a)The total offset voltage to be added to sinusoidal reference
signals which are added by the total offset voltage to signals (b)The A-phase sinusoidal reference signal after offset voltage
make SPWM scheme equivalent to the SVPWM scheme, is added (c) Motor phase voltage when M=0.1 (2-level operation)
are simultaneously compared with the triangular carrier
set. A DC link voltage (Edc) of 800 volts is assumed for
35
Figure 8. (a)The total offset voltage to be added to sinusoidal reference

Figure 6. (a)The total offset voltage to be added to sinusoidal reference signals (b)The A-phase sinusoidal reference signal after offset voltage
signals (b)The A-phase sinusoidal reference signal after offset voltage is added (c) Motor phase voltage when M=0.42 (5-level operation)
is added (c) Motor phase voltage when M=0.2 (3-level operation)
Figure 7. (a)The total offset voltage to be added to sinusoidal reference Figure 9. (a)The total offset voltage to be added to sinusoidal reference
signals (b)The A-phase sinusoidal reference signal after offset voltage signals (b)The A-phase sinusoidal reference signal after offset voltage
is added (c) Motor phase voltage when M=0.3 (4-level operation) is added (c) Motor phase voltage when M=0.53 (6-level operation)
36
Figure 10. (a)The total offset voltage to be added to sinusoidal reference Figure 12. (a)The total offset voltage to be added to sinusoidal reference
signals (b)The A-phase sinusoidal reference signal after offset voltage signals (b)The A-phase sinusoidal reference signal after offset voltage
is added (c) Motor phase voltage when M=0.62 (7-level operation) is added (c) Motor phase voltage when M=0.85 (9-level operation)
7. References
[1] Holtz, J., “Pulsewidth modulation–A survey,” IEEE
Trans. Ind. Electron., 30 (5), pp. 410–420,
1992.
[2] Zhou, K., and Wang, D., “Relationship between
space-vector modulation and three-phase
carrier-based PWM: A comprehensive
analysis,” IEEE Trans. Ind. Electron., 49 (1), pp.
186–196, 2002.
[3] Van der Broeck, Skudelny, H.C., and Stanke, G.V.,
“Analysis and realisation of a pulsewidth
modulator based on voltage space vectors,”
IEEE Trans. Ind. Appl., 24 (1), pp. 142–150, 1988.
[4] Boys, J.T., and Handley, P.G., “Harmonic analysis
of space vector modulated PWM waveforms,”
IEE Proc. Electr. Power Appl.,137 (4), pp. 197–204,
1990.
[5] Holmes, D.G., “The general relationship between
regular sampled pulse width modulation and
space vector modulation for hard switched
converters,” Conf. Rec. IEEE Industry Applications
Society(IAS) Annual Meeting, pp. 1002–1009, 1992.
[6] Kim, J., and Sul, S. “A novel voltage modulation
technique of the Space Vector PWM,” Proc.
Int. Power Electronics Conf., Yokohama,
Japan, pp. 742–747, 1995.
Figure 11. (a)The total offset voltage to be added to sinusoidal reference [7] Carrara, G.,Gardella, S.G., Archesoni,M., Salutari,
signals (b)The A-phase sinusoidal reference signal after offset voltage R., and Sciutto, G., “A new multi-level PWM
is added (c) Motor phase voltage when M=0.74 (8-level operation) method: A theoretical analysis,” IEEE Trans.
Power Electron., 7 (3), pp. 497–505, 1992
[8] K.Chandra Sekhar, G.Tulasi Ram Das, “An Eight-
level Inverter System for an Induction Motor with
Open-end Windings” IEEE PEDS 2005, pp:219-223,
2005.
[9] K.Chandra Sekhar, G.Tulasi Ram Das, “A Nine-
Level Inverter System for an Open-End Winding
37
Induction Motor Drive” in preceedings of IEEE [20]G.Sambasiva Rao, K.Chandra Sekhar “A novel nine-
ICIEA 2006. level inverter system for Dual- fed induction motor
[10]Shivakumar, E.G., Gopakumar, K., Sinha, S.K., drive” International Journal of Engineering Research
Andre, Pittet, and Ranganathan, V.T., “Space & Industrial Applications, 4 (III), pp: 159-176,
vector PWM control of dual inverter fed open- August 2011.
end winding induction motor drive,” Proc. Applied [21] P.Satish Kumar, J.Amarnath, S.V.L. Narasimham,
Power Electronics Conf. (APEC), pp. 339–405, “A Qualitative Space Vector PWM Algorithm for a
2001. Five-level Neutral Point Clamped Inverter” ICGST
[11]Suh, J., Choi, C., and Hyun, D., “A new simplified International Journal on Automatic Control and
space vector PWM method for three-level Systems Engineering, ACSE, 9 (I), pp:43-50, 2009.
inverter,” Proc. IEEE Applied Power Electronics [22]V. Naga Bhaskar Reddy and V. Narasimhulu and Dr.
Conf. (APEC), pp. 515–520, 1999. Ch. Sai Babu, “control of cascaded multilevel
[12]Baiju, M.R., Mohapatra, K.K., Somasekhar, V.T., inverter by using carrier based pwm technique and
Gopakumar, K., and Umanand, L. “A five-level implemented to induction motor drive” ICGST
inverter voltage space phasor generation for an International Journal on Automatic Control and
open-end winding induction motor drive,” IEE Systems Engineering, ACSE, 10 (II), pp:11-18,
Proc. Electr. Power Appl., 150 (5), pp. 531–538, December 2010.
2003.
[13]Wang, FEI, “Sine-triangle versus space vector Biographies
modulation for threelevel PWM voltage source G.Sambasiva Rao received B.E.
inverters,” Proc. IEEE-IAS Annual Meeting, Rome, degree in Electrical & Electronics
pp. 2482–2488, 2000. Engineering from Andhra
[14]Baiju, M.R., Gopakumar, K., Somasekhar, V.T., University Engineering College,
Mohapatra, K.K., and Umanand, L., “A space vector Visakha Patnam, India in 2000
based PWMmethod using only the instantaneous and M.E. with Power Electronics
amplitudes of reference phase voltages for three- & Industrial Drives from
level inverters,” EPE J., 13 (2), pp. 35–45, 2003. Sathyabama Institute of Science
[15]McGrath, B.P., Holmes, D.G., and Lipo, T.A., and Technology, Chennai, India in 2006. Since 2006, he
“Optimized space vector switching sequences for has been with R.V.R & J.C.College of engineering as
multilevel inverters,” Proc. IEEE Applied Power Lecturer. Presently he is a part-time research student at
Electronics Conf. (APEC), pp. 1123–1129, 2001. J.N.T.U College of Engineering, Hyderabad- 500072,
[16]Krah, J., and Holtz, J., “High performance current India, working towards his doctoral degree.
regulation and efficient PWM Implementation for
low inductance servo motors,” IEEE Trans. Ind. Dr.K.Chandra Sekhar received
Appl., 36 (5), pp. 1039–1049, 1999. his B.Tech degree in Electrical &
[17]Somasekhar, V.T., and Gopakumar, K., “Three-level
Electronics Engineering from
inverter configuration cascading two 2-level V.R.Siddartha Engineering
inverters,” IEE Proc. Electr. Power Appl., 150 (3), College, Vijayawada, India in
pp. 245–254, 2003. 1991 and M.Tech with Electrical
[18]R.S.Kanchan, M.R.Baiju, K.K.Mohapatra, Machines & Industrial Drives
P.P.Ouseph and K.Gopakumar, “Space vector PWM from Regional Engineering
signal generation for multilevel inverters using only College, Warangal, India in 1994. He Received the PhD
the sampled amplitudes of reference phase voltages,” degree from the J.N.T.U College of Engineering,
IEE Proc.-Electr. Power Appl., 152 (2), pp.297-309, Hyderabad- 500072, India in 2008. He is having 16 years
March 2005. of teaching experience. Currently he is professor and
[19]M.R.Baiju,K.Gopakumar,K.K.Mohapatra, V.T Head in the Department of EEE, R.V.R & J.C. College
Somasekhar and L.Umanand, “Five-Level inverter of Engineering, Guntur, India. His Research interests are
voltage-space vector phasor generation for an Open- Power Electronics, Industrial Drives & FACTS Devices.
End Winding Induction Motor Drive”. In: IEE
proc.-power Appl., 150 (5), pp.531-538, Sep2003.
38
Simple and Novel Generalized Scalar PWM Algorithm for Multilevel Inverter
Fed Direct Torque Controlled Induction Motor
C. Harikrihsna1, T. Bramhananda Reddy2, J. Amarnath3, and S. Kamakshaiah3

1
Mother Teresa Institute of Science and Technology, Sathupally, Andhra Pradesh, India
2
G. Pulla Reddy Engineering College (autonomous), Kurnool, Andhra Pradesh, India
3
JNTUH, Kukatpally, Hyderabad, Andhra Pradesh, India
chalasaniharikrishna@gmail.com, [tbnr, amarnathjinka@rediffmail.com], kamkshaiahsatram@yahoo.co.in
Abstract in torque, flux and currents are high in DTC. To reduce

This paper presents a simple and novel generalized PWM the steady state ripples, discrete space vector modulation
algorithm for n-level multilevel inverter fed direct torque (DSVM) algorithm has been proposed in [4]. Though,
controlled induction motor drives. The proposed DSVM algorithm reduces the torque ripple up to some
algorithm does not use the angle and sector information extent, it gives variable switching frequency operation of
and hence reduces the complexity involved in the PWM the inverter.
algorithm. The proposed algorithm has been developed To reduce the harmonic distortion and to obtain the
by using the concept of imaginary switching times, which constant switching frequency operation of the inverter,
are proportional to the instantaneous phase voltages only. nowadays many researchers have focused their interest
The proposed algorithm generates the modulating on pulsewidth modulation (PWM) algorithms. A detailed
waveforms for a n-level inverter, which will be compared survey on various PWM algorithms is carried out in [5]
with the level shifted triangular carriers to generate the and concluded that space vector pulsewidth modulation
switching pulses to the inverter. To validate the proposed (SVPWM) algorithm gives good performance. The
algorithm several numerical simulation studies have been generation of switching pulses in SVPWM algorithm is
carried out and results are presented. From the given in [6]. Though the SVPWM algorithm gives good
simulation results, it can be observed that the proposed performance, the complexity involved is more due to
algorithm gives reduced harmonic distortion with angle and sector calculations. To reduce the complexity
increased number of levels. involved in SVPWM algorithm, carrier based SVPWM
algorithm is developed in [7] by adding the offset voltage
Keywords: Space vector PWM, Multi Level Inverter DTC, to the reference phase voltages. Hence, to reduce the
Induction Motor. ripples and to obtain the constant switching frequency
operation of the inverter, SVPWM algorithm is used for
1. Introduction DTC in the literature [8]-[9].
All most the DC drives are replaced by AC drives The two-level inverter fed DTC drives are suitable for
because of its variable speed characteristics. The low power applications only. In order to meet the
invention of field oriented control (FOC) algorithm has medium and high power applications, nowadays
been made a renaissance in the high-performance multilevel inverters are becoming popular. A diode
variable speed drive applications. The FOC algorithm clamped three level inverter fed induction motor drive is
gives the decoupling control of torque and flux of an presented in [10]. A detailed survey on various types of
induction motor drive and control the induction motor multilevel inverters is given in [11]. But, as the number
similar to a separately excited dc motor [1]. But, the of levels increases, the complexity involved in the PWM
complexity involved in the FOC algorithm is more due to algorithms also increases. To reduce the complexity
reference frame transformations. To reduce the involved in the SVPWM algorithm, a simplified
complexity involved, a new control strategy called as SVPWM algorithm is presented for three-level inverter in
direct torque control (DTC) has been proposed in [2]. A [12].
detailed comparison is brought between FOC and DTC in To reduce the complexity, a novel voltage modulation
[3] and concluded that DTC gives fast torque response algorithm is presented in [13] by using the concept of
when compared with the FOC. Though, FOC and DTC effective time. By extending the same concept, various
give fast transient and decoupled control, these operate PWM algorithms have been generated for two-level
the inverter at variable switching frequency due to inverters in [14]. The concept of the PWM algorithms
hysteresis controllers. Moreover, the steady state ripples which is presented in [13]-[14] is extended for multilevel
inverters in [15].
39
 
This paper presents a simple generalized scalar PWM
algorithm for 2-, 3-, 5- and 7-level inverter. The proposed T2 
3
 Vd  * Ts  Ts  3 * Vd
Vdc Vdc
algorithm has been developed by using the concept of
imaginary switching times, which are proportional to the Ts  1 3    1 3 
instantaneous sampled phase voltages only. Moreover,   Vq  Vd 
 
Vq  Vd 

(5)
the proposed algorithm does not require the calculation of  2
Vdc 2   2 2 
angle and sector information and hence reduces the Ts Ts
complexity involved in the PWM algorithm.  Vbn  Vcn  Tbn  Tcn
Vdc Vdc
2. Simplified SVPWM Algorithm for Two-

level inverter
To reduce the complexity involved in the conventional
SVPWM algorithm and execution time due to longer
calculations, a simplified SVPWM algorithm has been
developed using the concept of imaginary switching
times. In this approach, the actual switching times for
each inverter leg are deduced directly as a simple form.
The proposed approach is based on the instantaneous
values of the reference voltages of a, b and c phases only.
This method does not depend on the magnitude of the
reference voltage space vector and its relative angle with
respect to the reference axis. The transformation from
two-phase voltages to three-phase voltages can be
obtained from the stationary frame reference voltages as
given in (1)
 
Van   1 0 
 3  Vq 
Figure. 1 Voltage space vectors produced by an inverter
V    1 (1)
 bn   2 2  Vd 
Vcn   1   From (4) – (5), the imaginary switching time periods
 3 
 2 2  proportional to the instantaneous values of the reference
phase voltages are defined as
If the reference voltage vector lies in the first sector as T 
shown in Figure 1, then the actual switching times can be Tan   s  Van
deduced as  Vdc 
3Vref Sin(60o   ) T 
T1  * Ts Tbn   s  Vbn (6)
2Vdc Sin60o  Vdc 
(2)
3  3  T 

1
Vref cos   Vref sin   * Ts Tcn   s  Vcn
Vdc  2 2 
  Vdc 
Hence, from (4) – (5), the active vector switching times
3Vref Sin
T2  * Ts T1 and T2, if the reference voltage vector falls in sector-1
2Vdc Sin60o may be expressed as
(3)
T1  Tan  Tbn ; T2  Tbn  Tcn (7)

3
Vdc
 
Vref cos  * Ts Thus, the active voltage vector switching times can be
represented by the time difference between the imaginary
But from Figure 1 it can be observed that switching time periods. In the SVPWM algorithm, when
Vq  Vref cos  and Vd   Vref sin  . Hence, the active the reference voltage vector falls in the first sector, the
vector times can be simplified as imaginary switching time which is proportional to the a-
3  3 1  phase ( Tan ) has a maximum value, the imaginary
T1  Vq  Vd  * Ts
Vdc  2 2  switching time which is proportional to the c-phase ( Tcn )
has a minimum value and the imaginary switching time
T  1 3   which is proportional to the b-phase ( Tbn ) is neither
 s Vq   Vq  Vd * Ts (4)
Vdc  2 2  minimum nor maximum switching time. Thus, in general
 
to calculate the active vector switching times, the
T T
 s Van  s Vbn  Tan  Tbn maximum and minimum values of imaginary switching
Vdc Vdc times are calculated in every sampling time as given in (8)
and – (9).
Tmax  Max(Tan , Tbn , Tcn ) (8)
Tmin  Min(Tan , Tbn , Tcn ) (9)
40
The effective time Teff can be defined as the time capacitors. Although each active switching device is only
difference between Tmax and Tmin and can be given as Vdc
required to block a voltage level of , the clamping
in (10). (n  1)
Teff  Tmax  Tmin (10) diodes must have different voltage ratings for reverse
voltage blocking. Assuming that each blocking diode
voltage rating is the same as the active device voltage
The effective time means the duration in which the rating, the number of diodes required for each phase will
effective voltage is supplied to the machine terminals. In be (n  1)  (n  2) . This number represents a quadratic
the actual switching instants, there is one degree of
freedom that the effective time can be located anywhere increase in n. when n is sufficiently high, the number of
within one sampling interval. To generate actual diodes required will make the system impractical to
switching pattern which preserves the effective time, the implement. If the inverter runs under PWM, the diode
zero sequence time is subjoined to the phase voltage time. reverse recovery of these clamping diodes becomes the
In order to locate the effective time in centre of the major design challenge in high-voltage high-power
sampling interval, the zero sequence voltage has to be applications.
symmetrically distributed at the beginning and end of one
sampling period. Therefore, the actual switching times 4. Proposed Generalized PWM Algorithm
for each inverter leg can be simply obtained by the time for an n-Level Inverter
shifting operation as below. For a 2-level inverter only one offset time is sufficient to
Tga  Tas  Toffset evaluate the switching instants. Because, in a two-level
SVPWM algorithm, in a sampling time interval, the
Tgb  Tbs  Toffset (11)
imaginary switching time which has lowest magnitude
Tgc  Tcs  Toffset ( Tmin ) crosses the triangular carrier first, and causes the
first transition in the inverter switching state. While the
To distribute zero voltage symmetrically during one imaginary switching time, which has the maximum
sampling period, the offset time Toffset is achieved using magnitude ( Tmax ), crosses the carrier last and causes the
a simple sorting algorithm. the zero voltage vector time last switching transition in the inverter switching states.
duration can be calculated as given in (12). Thus the switching periods of the active vectors can be
Tzero  Ts  Teff (12) determined from the Tmax and Tmin in a two-level
inverter scheme.
And, Tmin  Toffset  Tzero / 2 (13)
But, the SVPWM algorithm for multilevel inverters,
Therefore, Toffset  Tzero / 2  Tmin (14) involves comparing the modulating signals with a
number of symmetrical level-shifted triangular carrier
In order to generate symmetrical switching pulse pattern waves for PWM generation. It has been shown that for an
within two sampling intervals, when the switching n-level inverter, n-1 level-shifted triangular carrier waves
sequence is ‘ON’ sequence, the actual switching time are required for comparison with the modulating
will be replaced by the subtraction value with the waveforms. Because of the level-shifted multicarriers, the
sampling time as follows: first crossing (termed the first-cross) of the modulating
wave cannot always be the min-phase. Similarly, the last
crossing (termed the third-cross) of the modulating wave
Tga , gb, gc  Ts  Tga, gb, gc (15)
cannot always be the max-phase. Thus, a single offset
time is not sufficient to centre the middle inverter
As described above, the effective time implies the switching vectors, in a multilevel PWM scheme during a
applied time of a certain active vector. Therefore, with sampling period Ts.
the effective vector concept, the actual switching time The proposed algorithm determines the offset time for the
can be obtained directly from the stationary frame PWM generation in the linear modulation region by using
reference voltage without sector identification, effective the concept of imaginary switching times. The proposed
time calculation and recombination. generalized PWM algorithm presents a simple way to
To implement the novel space vector PWM, only a 3- determine the time instants at which the three reference
element simple sorting algorithm is needed. Therefore, phases cross the triangular carriers. These time instants
the calculation efforts of the proposed PWM method is are sorted to find the offset voltage to be added to the
greatly reduced as compared with the conventional space reference phase voltages for SVPWM generation for
vector PWM method. multilevel inverters for the entire linear modulation range,
so that the middle inverter switching vectors are centered
3. Topology of n-level Diode-Clamped (during a sampling interval), as in the case of the
Multilevel Inverter conventional two-level SVPWM scheme.
An inverter which is capable of three-level or more is For an n-level inverter, the instantaneous imaginary
termed as Multilevel Inverter. The topology of n-level switching times are calculated as follows:
diode clamped multilevel inverter is as shown in Figure. Ts
Tas  Van  (16)
2. In this circuit, the dc-bus voltage will be split into Vdc /(n  1)
suitable number of levels according to the number of
41
Ts The time equivalent of a common mode voltage is termed

Tbs  Vbn  (17) as Toffset1 which can be defined as
Vdc /( n  1)
Ts Toffset1  (Tmax  Tmin ) / 2 (19)
Tcs  Vcn  (18)
Vdc /(n  1) where Tmax and Tmin are the maximum and minimum
where n represents number of levels. of Tas , Tbs and Tcs .
Cn
Cn-1
C4
C3
C2 C3
V0
C2
C1
C2
C1
C1
2 level line
C1
C2
3 level line
4 level line
5 level line C1
n level line
Figure. 2 Topology of n-level multilevel inverter
Then time equivalents of the modified phase voltage Tc  cross when n is odd, can be determined for a n-level
magnitudes can be defined as inverter as:
*
Tas  Tas  Toffset1 (20) *
Ta _ cross  Tas  (( I a  (n  1) / 2) * Ts (23)
*
Tbs  Tbs  Toffset1 (21) *
Tb _ cross  I bs  (( I b  (n  1) / 2) * Ts (24)
Tcs*  Tcs  Toffset1 (22) *
Tc _ cross  Tcs  (( I c  (n  1) / 2) * Ts (25)
The triangular carriers and the modulating waves, for an When n is even, the time durations, Ta  cross , Tb  cross
n-level PWM scheme are shown in Figure. 3a, for n is
odd, and in Figure. 3b, for n is even. The (n-1) triangular and Tc  cross , can be determined for a n-level inverter as:
carriers are compared with reference phase voltages as Ta _ cross  Ts / 2   Tas
*
 (( I a  (n / 2)) * Ts ) (26)
shown in Figures. 3a and 3b. A carrier index, I, is defined
to designate the carrier regions in which the reference Tb _ cross  Ts / 2   Tbs
*
 (( I b  (n / 2)) * Ts ) (27)
phase voltages lie during the sampling interval under
consideration. During a sampling interval, the carrier Tc _ cross  Ts / 2   Tcs*  (( I c  (n / 2)) * Ts ) (28)
indices, Ia, Ib and Ic (which can be from 1 to n-1), for a, b in the proposed generalized PWM algorithm, the
and c phases, respectively, are determined depending on Ta  cross , Tb  cross and Tc  cross are used to centre the
the carrier region in which the respective phase voltage
middle switching vectors as in the case of two-level
lies. Then the time durations, Ta  cross , Tb  cross and inverters in each sampling time interval. The time-
duration, at which the reference phases cross the
42
triangular carriers for the first time, second time and third By adding the time, Toffset 2 to T first _ cross ,
time are defined as T first _ cross , Tsec ond _ cross and
Tsec ond _ cross and Tthird _ cross gives the inverter leg
Tthird _ cross respectively, in a sampling time interval Ts.
switching times Tga , Tgb and Tgc for phases a, b and c,
these can be calculated from (29).
 
T first _ cross  min T x _ cross
respectively as (34)-(36).
Tga  Ta _ cross  Toffset 2 (34)

Tsec ond _ cross  mid T x _ cross  (29)
Tgb  Tb _ cross  Toffset 2 (35)
 
Tthird _ cross  max T x _ cross ; x  a, b, c
Tgc  Tc _ cross  Toffset 2 (36)
A B C Vdc
1 2
5. Results and Discussion
2 To validate the proposed generalized scalar PWM
algorithm, several numerical simulation studies have
been carried out and results are presented. The details of
the induction motor, which is used for simulation studies
are as follows:
0
A 3-phase, 4 pole, 4kW, 1200rpm induction motor with
parameters of Rs = 1.57Ω, Rr = 1.21Ω, Ls = Lr = 0.17H,
Lm = 0.165H and J = 0.089Kg.m2.
The steady state simulation results for 2-, 3-, 5-, and 7-
(n-2) level inverter fed DTC-IM drive are shown from Figure.
4 to Figure. 19. From the simulation results it can be
(n-1)  Vdc
2
observed that as the number of levels increases the
(a) harmonic distortion will be decreased. Also, it can be
A B C concluded that the proposed PWM algorithm is efficient
Vdc with reduced complexity. The total harmonic distortion
1 2 (THD) values of line currents and voltages are given in
Table-1.
2 TABLE I
COMPARISON OF THD IN LINE CURRENT FOR VARIOUS LEVELS
Level % THD in line
current % THD in line voltage
0
2 level 9.84 75.96
3 level 4.06 38.10
5 level 2.23 17.25
7 level 1.37 12.17
(n-2)
(n-1)  Vdc
2
(b)
Figure. 3 Triangular carrier and reference signals (a) n-level PWM
scheme where n is even (b) n-level PWM scheme where n is even
During one sampling interval Ts, the time durations

T first _ cross , Tsec ond _ cross and Tthird _ cross directly
decides the switching times for different inverter voltage
vectors. An offset time, Toffset 2 is added to T first _ cross ,
Tsec ond _ cross and Tthird _ cross to centre the middle
vector. The offset time, Toffset 2 is determined as follows:
The time duration for the middle inverter switching
vectors, Tmiddle , is given as
Tmiddle  Tthird _ cross  T first _ cross (30)
The time duration for start and end vector is
T0  Ts  Tmiddle (31)
The time duration for start vector is given by
T0 / 2  T first _ cross  Toffset 2 (32)
Therefore, Toffset 2  T0 / 2  T first _ cross (33) Figure. 4 Simulation results of 2-level SVPWM based DTC: steady-
state plots of speed, torque, currents and flux at 1200 rpm
43
Figure. 10 Harmonic Spectrum of line current for 3-level SVPWM

based DTC along with THD.
Figure. 5 Simulation results of 2-level SVPWM based DTC: steady
state plots of phase and line voltages
Figure. 11 Harmonic Spectrum of line voltage for 3-level SVPWM

based DTC along with THD.
based DTC along with THD
Figure. 7 Harmonic Spectrum of line voltage for 2-level SVPWM based

DTC along with THD
Figure. 12 Simulation results of 5-level SVPWM based DTC: steady-

Figure.8 Simulation results of 3-level SVPWM based DTC: steady-state

plots of speed, torque, currents and flux at 1200 rpm Figure. 13 Simulation results of 5-level SVPWM based DTC: steady

Figure.9 Simulation results of 3-level SVPWM based DTC: steady based DTC along with THD
44
6. Conclusion
A simple and novel generalized SVPWM algorithm is
presented for an n-level inverter fed direct torque
controlled induction motor drive. The proposed algorithm
has been developed by using the concept of imaginary
switching times, which are proportional to the sampled
reference phase voltages only. Hence, the complexity
involved id less when compared with the classical PWM
Figure. 15 Harmonic Spectrum of line voltage for 5-level SVPWM algorithms. To validate the proposed algorithm, the
based DTC along with THD simulation studies have been carried out and results are
presented. From the simulation results, it can be
concluded that the proposed algorithm gives good
performance with reduced complexity. Moreover, as the
number of levels increases, the harmonic distortion of
line current and voltages also decreases.
7. References
[1] F. Blaschke “The principle of field orientation as
applied to the new transvector closed loop control
system for rotating-field machines," Siemens
Review, 1972, pp 217-220.
[2] Isao Takahashi and Toshihiko Noguchi, “A new
quick-response and high-efficiency control strategy
of an induction motor,” IEEE Trans. Ind. Applicat.,
vol. IA-22, no.5, Sep/Oct 1986, pp. 820-827.
[3] Domenico Casadei, Francesco Profumo, Giovanni
Serra, and Angelo Tani, “FOC and DTC: Two
Figure. 16 Simulation results of 7-level SVPWM based DTC: steady- Viable Schemes for Induction Motors Torque
Control” IEEE Trans. Power Electron., vol. 17, no.5,
Sep, 2002, pp. 779-787.
[4] Domenico Casadei, Francesco Profumo, Giovanni
Serra, and Angelo Tani, “FOC and DTC: Two
Viable Schemes for Induction Motors Torque
Control” IEEE Trans. Power Electron., vol. 17, no.5,
Sep, 2002, pp. 779-787.
[5] Joachim Holtz, “Pulsewidth modulation – A survey”
IEEE Trans. Ind. Electron.., vol. 39, no. 5, Dec 1992,
pp. 410-420.
Figure. 17 Simulation results of 7-level SVPWM based DTC: steady [6] Heinz Willi Vander Broeck, Hnas-Christoph
Skudelny and Georg Viktor Stanke, “Analysis and
realization of a pulsewidth modulator based on
voltage space vectors” IEEE Trans. Ind. Applicat.,
vol. 24, no. 1, Jan/Feb 1988, pp. 142-150.
[7] Ahmet M. Hava, Russel J. Kerkman and Thomas A.
Lipo, “Simple analytical and graphical methods for
carrier-based PWM-VSI drives” IEEE Trans. Power
Electron., vol. 14, no. 1, Jan 1999, pp. 49-61.
based DTC along with THD [8] T. Brahmananda Reddy, J. Amarnath and D.
Subbarayudu, “Direct torque control of induction
motor based on Hybrid PWM method for reduced
ripple: A sliding mode control approach” ACSE
Journal Vol.(6) Issue(4) Dec.2006,pp.23-30.
[9] Lixin Tang and M.F. Rahman, “A new direct torque
control strategy for flux and torque ripple reduction
for induction motors drive by using space vector
modulation” in Conf. Rec. of IEEE-PESC, 2001,
Figure. 19 Harmonic Spectrum of line voltage for 7-level SVPWM pp.1440-1445.
based DTC along with THD
45
[10] Nabae, A., Takahashi, I., and Akagi, H, "A neutral- Biographies
point clamped PWM inverter’, IEEE-Trans. Ind.
Chalasani Hari Krishna was born in 1982.
Appl., 1981, 17, (5), pp.518-523. He graduated from Jawaharlal Nehre
[11] Jose Rodriguez, Jih-Sheng Lai, Fang Zheng Peng: Technological University, Hyderabad in the
year 2003. He received M.E degree from
‘Multilevel Inverters: A Survey of Topologies, Satyabama University, Chennai in the year
Control, and Applications’, IEEE Trans. On Ind. 2005. He presently Associate Professor in
Elec., Vol. 49, No. 4, Aug. 2002. the Department of Electrical and Electronics
Engineering at Mother Teresa Institute of
[12] Abdul Rahiman Beig, G. Narayana, V.T. Science and Technology, India. His research
Ranganathan, “Modified SVPWM Algorithm for area includes DTC and Drives.
Three Level VSI With Synchronized and
Symmetrical Waveforms”, IEEE Trans. Ind. Elect.,
Dr. T. Brahmananda Reddy was bornin
Vol. 54, No. 1, Feb. 2007, pp 486-494. 1979. He graduated from Sri Krishna
[13] Joohn-Sheok Kim and Seung-Ki Sul, “A novel Devaraya University, Anantapur in the year
voltage modulation technique of the space vector 2001. He received M.E degree from
Osmania University, Hyderabad, India in the
PWM” in Proc. IPEC, Yokohama, Japan, 1995, pp. year 2003 and Ph.D from J.N.T.University,
742-747. Hyderabad in the year 2009. He is
presently Associate Professor in the
[14] Dae-Woong Chung, Joohn-Sheok Kim and Seung-Ki Electrical and Electronics Engineering
Sul, “Unified voltage modulation technique for real- Department, G. Pulla Reddy Engineering
College, Kurnool, India. He presented more
time three-phase power conversion” IEEE Trans. than 90 research papers in various national and international
Ind. Applicat., vol. 34, no. 2, Mar/Apr 1998, pp. conferences and journals. His research areas include PWM techniques,
DC to AC converters and control of electrical drives
374-380.
[15] R.S. Kanchan, M.R. Baiju, K.K. Mohapatra, P.P.
Ouseph and K. Gopakumar, "Space vector PWM Dr. J. Amarnath graduated from Osmania
University in the year 1982, M.E from
signal generation for multilevel inverters using only Andhra University in the year 1984 and
the sampled amplitudes of reference phase voltages" Ph.D from J.N.T.University, Hyderabad in
the year 2001.He is presently Professor and
IEE Proc.-Electr. Power Appl., vol. 152, No. 2, pp. Head of the Electrical and Electronics
297-309, March, 2005. Engineering Department, JNTU College of
Engineering, Hyderabad and also he is the
Chairman, Board of studies in Electrical
and Electronics Engineering, JNTU
College of Engineering, Hyderabad. He presented more than 100
research papers in various national and international conferences and
journals. His research areas include Gas Insulated Substations, High
Voltage Engineering, Power Systems and Electrical Drives.
Dr. S. KAMAKSHAIAH is from JNT

University, Hyderabad and worked as
Professor, Head of the Department &
Chairman of Electrical Sciences. He
annexed Gold Medal for getting 1st rank in
SV University at B.Tech level. He was also
awarded Gold Medal for Ph.D. work from
I.I.Sc.Bangalore for its industrial application.
He was the Best Teacher award winner of
the year 1997-98 which was promoted by
A.P.State Government. He has got 40
Technical Publications in International level. He is also an author of
three text books Introduction to Electrical Engineering, Basic Electrical
Engineering and Electrical Technology with his co – author Prof. Naidu,
and HVDC Transmission published by Tata McGraw Hill. He carried
research work in the area of Vaccum arcs in Triggered Vaccum Gaps
(TVG). At present he is working on the problems of Gas Insulated
Substation (GIS) ,HVDC and Electric Drives. Presently he is working
as Director, Krishna Chaitanya Institute of Technology & Science,
Markapur, He visited various countries like Canada, Germany and
Doha and presented papers at IEEE International Conference at
ATLANTA, USA in the year June 2010.
46
Amélioration du temps de convergence de l’algorithme

de calcul du filtre optimal par HOS en DMT
H. Bellahsene, I. F. E. Fatani
Département TCSN, Faculté SNV, Université de A/Mira Béjaia 06000, Algérie
[bellahsene_ho]@yahoo.fr,
http://www.univ-bejaia.dz
Abstract optimales et/ou adaptatives [12] réalisent le même

travail avec moins de performances. La plupart
Ce travail synthétise l’évolution de l’égalisation des approches par modèle TEQ (Time domaine
classique vers l’égalisation par réduction de canal. Il Equalizer) ne sont pas adaptatives, elles nécessitent
s’agit d’utiliser les statistique d’ordre supérieure à l’apprentissage ou une estimation du canal avec des
deux à savoir le calcul du maximum de kurtosis pour calculs assez lourds à mettre en œuvre. L’algorithme
retrouver le filtre TEQ (Time-domain EQualizer) op- amélioré est celui de MERRY (Multicarrier Equali-
timal. Le délai de synchronisation dans l’algorithme sation by Restoration of RedundancY) [6, 11]. Nous
du MSSNR (Maximum Shortening Signal to Noise proposons une méthode basée sur les HOS, dite par
Ratio) est retrouvé par la même technique. Une maximum du kurtosis, et nous calculons également le
comparaison entre le filtre optimal retrouvé par délai optimum ainsi que le filtre optimal lui correspon-
d’autres méthodes et celui déterminé par la méthode dant.
la notre. La fonction statistique gamma a réduire de Cet article est organisé de la façon suivante: Section
manière conséquente le délai de convergence après 2 met en évidence l’algorithme d’égalisation multi-
son introduction dans l’algorithme.1 porteuse par rétablissement de redondance.La fonc-
tion coût sera donnée en Section 3. Notre contribu-
Keywords: TEQ (Time-domaine EQalizer), tion sera mentionnée en Section 4. Des simulations
High Order Statistics, kurtosis, DMT, fonction et discutions des résultats fairront l’objet de la partie
Gamma. théorique en Section 5.Les résultats importants ainsi
que les futurs travaux envisagés seront donnés en con-
clusion.
1 Introduction
L’évolution de l’égalisation classique vers l’égalisation
par réduction de canal à fait l’objet de nombreux
2 Algorithme d’égalisation mul-
travaux dans la litérature. Nous nous sommes in- tiporteuse par rétablissement
téressés à l’introduction des HOS dans ce damaine.
Nous appliquons le calcul du maximum de kurtosis
de redondance
pour retrouver le filtre TEQ (Time-domain EQual-
L’algorithme MERRY (Multicarrier Equalization by
izer) optimal ainsi que le délai de synchronisation dans
Restoration of RedundancY) exploite la redondance
l’algorithme du MSSNR (Maximum Shortening Sig-
introduite dans les données à travers le préfixe cy-
nal to Noise Ratio). Nous comparons le filtre optimal
clique, voir figure (2)[6]. C’est une approche basée
retrouvé par d’autres méthodes et celui déterminé par
sur la minimisation de l’espérance mathématique du
la méthode du maximum du kurtosis. Nous arrivons
carré de la différence cyclique qui est la différence en-
par utilisation des fonctions statistiques notamment
tre des valeurs de séparées par un bloc de données.
gamma, à réduire de manière conséquente le délai de
convergence au prix de la stabilité de l’algorithme.
Des méthodes retrouvent le canal et calculent
y(2) = h0 s(2) + h1 s(1) + h2 s(0) + h3 s(−1) + h4 s(−2)
l’égaliseur optimal [8]. D’autres sont dites sous-
= h0 s(10) + h1 s(9) + [h2 s(0) + h3 s(−1) + h4 s(−2)]
1 Ce travail à été réalisé au département des TCSN, FSNV. y(10) = h0 s(10) + h1 s(9) + [h2 s(8) + h3 s(7) + h4 s(6)] (1)
47
avec δ ∈ {0, · · · , M − 1}, δ étant le délais désiré (c’est

le coefficient de synchronisation). Le gradient stochas-
tique de (5) donne l’algorithme de réduction de canal
adaptatif aveugle MERRY (Multicarrier Equalization
by Restoration of RedundancY) [MAR 05], avec: Pour
un δ donné, le symbole k = 0, 1, 2, · · · ,
Figure 1: Rajout du CP à la trame de données r̃k) = r(M k + P + δ) − r(M k + P + N + δ)

e(k) = f T (k)r̃k)
La figure 2 fait appel à l’équation 1. Ainsi, f̂(k + 1) = f T (k) − µe(k)r̃k)

les N + P données transmises deviennent péri- f̂(k + 1)
f(k + 1) = (6)
odiques. Cette périodicité est nécessaire pour main-
f̂(k + 1)
tenir l’orthogonalité des multiporteuses [10]. Les
derniers symboles sont identiques aux premiers et T ( )
nous pouvons écrire : avec r(i) = [r(i), r(i − 1), · · · , r(i − Lf ] , et ∗ est
le conjugué complexe. Pour éviter la solution triviale
s(M k + i) = s(M k + i + N ), i ∈ {1, · · · , N } (2) (f = 0 ) nous devons prendre ∥f∥ = 1. L’algorithme
représenté par (6) donne
[ ] le vecteur propre minimal de
où M = P + N est la durée de tous les symboles et la matrice A = E r̃, r̃H . Soit C la matrice canal de
k est l’indice de chaque symbole. La figure 2 montre convolution autrement dit h = Cf et soit Cwall obtenue
un exemple où l’on prend N = 8, P = 2 et M = par suppression de δ lignes des δ + P − 1 de C.
10 correspondant à k = 0. Les données reçues sont Si l’entrée s(k) est blanche, alors A = CTwall , Cwall .
obtenues à partir de selon la relation suivante : L’énergie à l’extérieur de la fenêtre du canal effectif
plus le gain du bruit est donnée comme suit :
∑
Lc
 
r(M k + i) = cl · s(M k + i − l) + n(M k + i) (3) ∑
δ−1 ∑Lh
Jδ = 2σs2  |hj |  + 2f T Rn f ∗ (7)
2 2
l=0 |hj | +
j=0 j=δ+P
Les données égalisées sont obtenues à partir de suivant
cette équation : Dans le cas où la réduction du canal utilisée est au-
todidacte
[ ] et non adaptative, alors la matrice A =
∑
Lf
E r̃, r̃H peut être estimée à partir des données. Le
y(M k + i) = fj · r(M k + i − j) (4)
vecteur propre correspondant à la valeur propre min-
j=0
imale peut être calculé, ce qui fournira une technique
La relation donnée par (1) est détériorée à cause du d’initialisation pour éviter les modes de convergence
canal lors de la réception des données, car les ISI qui lents.
affectent le CP sont différents de ceux qui affectent Réduction optimale du canal
les P derniers éléments du symbole. Considérons La réduction de canal optimale utilise un algorithme
l’exemple de la figure , Les éléments transmis 2 et 10 basé sur le calcul des valeurs propres et des vecteurs
sont identiques. Cependant, au niveau du récepteur, propres pour générer les coefficients du filtre réduit
l’interférence des éléments avant l’élément 2 ne sont optimal. La technique consiste à déterminer le filtre
pas toutes égales à celles qui sont avant l’élément 10. TEQ qui maximise le rapport de réduction du signal
Si on considère h2 , h3 et h4 nuls, alors y(2) = y(10) car sur le bruit dans une fenêtre. Pour ce faire, il faut
s(2) = s(10) et s(1) = s(9). Il est clair que pour avoir minimiser l’énergie de la réponse impulsionnelle du
il faut forcer h2 , h3 et h4 à zéro (éléments se trouvant canal effectif, à l’extérieur de la fenêtre contenant les
entre crochets dans l’équation 1), cette égalité étant (P +1 ) échantillons. Nous reprenons Cwin et Cwall tels
prise dans le sens des moindres carrés. Nous rendons que définies dans [8], alors hwin = Cwin f représente le
de cette façon le canal effectif de taille équivalente à canal effectif situé dans la fenêtre de taille ( P + 1).
celle du CP. Cependant hwall = Cwall f donne le canal effectif en
dehors de cette fenêtre. Le problème de conception
du TEQ à base du MSSNR peut être décrit comme
3 Algorithme MERRY et sa suit[8, 1, 2, 4, 9]:
( )
fonction coût min f T Af sous la contrainte f T Bf = 1 (8)
f
La fonction coût de l’algorithme MERRY est :
où A et B sont des matrices réelles [MAR 02],
[ ] symétriques Lf × Lf avec :
2
J = E |y (M k + P + δ) − y (M k + P + +N + δ)|
(5) A = CTwall , Cwall , B = CTwin , Cwin (9)
48
Elles représentent, respectivement, les énergies corre- Réduire le canal revient à trouver un filtre tel que sa
spondant à Cwall et Cwin où : séquence de sortie ait la même variance et le même
kurtosis maximum en valeur absolue qu’à l’entrée du
hTwall hwall = f T CTwall Cf = f T Af (10) filtre. Le filtre optimal sera celui qui maximisera la
quantité donnée par l’équation (19). Le délai de syn-
hTwin hwin = f T CTwin Cwin f = f T Bf (11)
chronisation sera la valeur de l’indice qui correspondra
La résolution de l’équation (8) mène au TEQ qui à ce filtre, tel que:
donne la solution du problème de vecteurs propres
généralisé [10], nous avons par conséquent : k = arg {max(gn )} (18)
n
Bf = λAf (12)
4.2 Calcul du pas de convergence par
où λ est la valeur propre généralisée correspondant kurtosis
au vecteur propre généralisé f, qui est la solution de
l’équation (12). Le kurtosis normalisé d’une variable aléatoire s ayant
comme valeur moyenne s̄ est défini par le rapport du
(√ )−1 (√ )−1
moment d’ordre quatre η4 au carré du moment d’ordre
C= BT A B (13)
deux η22 tous deux centrés. Les HOS (Higher Order
Statistics) définissent le kurtosis comme suit [LAC 97]:
L’inverse de la matrice A est retrouvé par factorisation
de Cholesky. Les TEQ sont alors donnés par : η4
κ4 (s) = (19)
(√ )−1 η22
f= BT q (14)
où, si est E l’opérateur espérance mathématique:
où q est le vecteur propre correspondant à la valeur { }
2
η2 = E (s − s̄) (20)
propre λ. Il est montré aussi [8] que l’algorithme
converge globalement si q = qmin est le vecteur pro- { }
4
pre minimal correspondant à la valeur propre mini- η4 = E (s − s̄) (21)
male λmin . C’est ainsi que le filtre TEQ optimal
L’excès (kurtosis standard) est défini par l’équation
fopt est retrouvé [3]. L’expression du SNR réduit opti-
(19) moins trois. Il s’annule lorsqu’il s’agit d’une dis-
mal (Shortening SNR ou SSNR) est donnée par :
tribution normale [3, 5]. D’ou tout l’intérêt à intro-
( ) ( )
T
fopt Bfopt 1 duire ce coefficient. En effet, sachant que le bruit ad-
SSN Ropt = 10 log T = 10 log ditif est dans la plupart des cas de type Gaussien, il
fopt Afopt λmin
est bien connu que les statistiques d’ordre supérieur ne
SSN Ropt = −10 log (λmin ) (15) sont pas affectées par le bruit. Nous proposons dans
ce papier une méthode itérative pour le calcul de µ. En
En somme, l’algorithme optimal utilise la matrice C, effet, le pas de convergence divisé par le kurtosis nor-
dont les éléments sont calculés à partir de Hwall et malisé donné par l’équation (19) de la séquence reçue à
Hwin . Le SSNR optimal est retrouvé par calcul direct l’entrée de l’égaliseur permet de tenir compte à chaque
des valeurs propres minimales de la matrice C. Les itération de la séquence à l’entrée de l’égaliseur r(i).
coefficients du TEQ optimal sont finalement calculés Ainsi, le pas de convergence du nouvel algorithme est
par transformation linéaire du vecteur propre unitaire calculé à partir d’un µ0 divisé à chaque itération par
associé à la valeur propre minimale. le kurtosis de sorte à avoir:
µ0
µi+1 = (22)
4 Modification de l’algorithme κ4 (r(i))
Par ce procédé, la mise à jour du pas de convergence
4.1 Le filtre optimal dépend des données reçues (c’est à dire des conditions
Cette procédure débute à partir de tous les filtres ré- de transmission du canal). Deplus, nous utilisons
duits fn ayant été trouvés à partir de l’équation (14), la fonction de densité de probabilité Gamma connue
utilisant l’algorithme du MSSNR [8]. Nous proposons dans la littérature pour la détermination du pas de
de choisir le filtre SCE (Shortened Channel Equalizer) convergence. Les résultats obtenus vont maintenant
optimal, non pas en comparant les énergies de tous les être discutés dans la partie consacrée aux simulations.
filtres trouvés tels que abordés par Martin et al. dans
[7], mais plutôt en calculant le kurtosis de chaque fil-
tre. Nous avons ainsi : 5 Simulations
yn = h ∗ fn (16) Les données que nous utilisons lors de la simulation
sont les huit CSA loop (Carrier Serving Area) [7], util-
gn = κ4 (yn) (17) isées dans le standard ADSL. Nous utilisons le CSA1
49
lors de nos simulations. Le CP est de taille, la taille
Fonction Coût MERRY

0
10
de l’égaliseur est égale à 16 [9]. La figure 2 représente ADAPTEE
MERRY OPTIMAL
les réponses impulsionnelles du canal réelle et celle
réduite par MSSNR à partir de l’algorithme MERRY
optimal. La figure3 représente les réponses impulsion-
0 100 200 300 400 500 600 700 800 900 1000
1 x 10
6
Débit binaire (bps)

3
Amplitudes normalisées
0.5
2
Adaptée
1 MERRY Optimal
MAX SSNR
0 0
0 100 200 300 400 500 600 700 800 900 1000
−0.5
Figure 4: Courbe de convergence par MERRY optimal

−1
0 50 100 150 200
Fonction Coût MERRY

Nombre de coefficients 0
10
ADAPTEE et KURTOSIS
MERRY OPTIMAL
Figure 2: Canal réel et canal réduit par MSSNR
nelles du canal réelle et celle réduite par la méthode

du maximum du kurtosis. Afin de pouvoir comparer 0 100 200 300 400 500 600 700 800 900 1000
les résultats donnés par les deux méthodes, nous nous x 10

6
4
débit binaire (bps)
proposons de dérouler l’algorithme de MERRY et

celui que nous proposons. Le pas de convergence étant 3
ADAPTEE et KURTOSIS
1 MERRY OPTIMAL
1 MAX SSNR
0
0 100 200 300 400 500 600 700 800 900 1000
indice symbole
Amplitudes normalisées
0.5
Figure 5: Courbe de convergence par kurtosis et la

0
fonction Gamma
−0.5
Il est à remarquer que les délais optimaux sont
différents de ceux retrouvés dans la littérature [6].
−1
0 50 100 150 200
Nombre de coefficients
Figure 3: Canal réel et canal réduit par maximum du

6 Conclusion
kurtosis et la fonction gamma Nous avons amélioré les approches adaptatives de ré-
duction de canal aveugle. En effet, une nouvelle méth-
déterminé par notre technique d’itération dépendant ode de calcul du filtre optimal par maximisation du
de la séquence de données ou de façon classique. Dans kurtosis à été proposée. Le délai de synchronisation
tous les cas nous choisissons un SNR égal à 40dB. a pu être calculé. Les résultats trouvés ont été com-
Les figures (2 et 3) donnent les réponses impul- parés avec ceux donnés dans la littérature. A travers
sionnelles réelle et réduite dans les deux cas (à savoir les HOS (High Order Statistics) conjugué à la fonction
la méthode du MSSNR et celle du kurtosis conjuguée Gamma le temps de convergence de l’algorithme a été
à la fonction Gamma. Les deux figures sont très sem- nettement réduit. Il est à signaler que la stabilité de
blables mis à pars quelques légères différences. La la fonction coût et du débit est meilleure avec la méth-
figure 4 donne la fonction coût de MERRY ainsi que ode de MERRY. La stabilité des résultats fera l’objet
le débit. Nous remarquons que la convergence est ef- de travaux futurs. Tous nos travaux de simulation ont
fective à partir de la 500ème itération avec un débit été réalisés sous Matlab.
fluctuant entre 3 et 4Mb/s. La figure 5 représente la
fonction coût et le débit faisant appel aux statistiques
d’ordre supérieur HOS et la fonction Gamma. Nous Remerciements
remarquons que la convergence est effective à partir
de la 180ème itération avec un débit fluctuant entre 3 Nous remercions le Pr J. M. Rouvaen du laboratoire
et 4Mb/s. IEMN de l’université de valenciennes en France pour
50
son aide. Nous remercions aussi, l’université de Texas Biographies

AT, Austin aux USA pour avoir mis à notre disposi-
tion les données standards CSA. Dr.Bellahsene Hocine a reçu
son diplôme d’ingénieur en communica-
tion en 1993 et le diplôme de Magistère
References 1999 à l’université de Tizi-ouzou
(Algérie). Il prépara sont diplôme de
[1] N. Al-Dhahir. FIR Channel Shortening Equal- doctorat au laboratoire IEMN labora-
izers for MIMO ISI Channels. IEEE Trans. on tory (Valenciennes, France) de janvier
comm, 49(2):213–218, 2001. 2005 à juillet 2007. Il reçu son doctorat
en juin 2009. Il est actuellement maître de conférence à
[2] G. Arslan, B. Lu, L. D. Clark, and B. L. l’université de Béjaia (Algérie). Son domaine d’intérêt
Evans. Iterative Refinement Methods for Time porte sur l’égalisation aveugle dans les systèmes de
Domain Equalizer Design. EURASIP Processing, coommunications, les statistiques d’ordre supérieur et les
2006(1):83–95, 2005. techniques MIMO pour les communications numériques.
mail: bellahsene_ho@yahoo.fr.
[3] I. Chahed, H. Bellahsene, J. M. Rouvaen, and
M. Djeddi. Selection of the order of ARMA
model using Kurtosis minimisation. Application Fatani Imade Fahd Eddine est
actuellement en post doc à l’univeristé
to blind deconvolution of seismic data. AMSE
de Valenciennes en France. Il à obtenu
proceding, 47(2):77–89, 2004. son diplôme de docteur en télécom-
[4] D. D. Falconer, S. L. Ariyavisitaul, A. Benyamin- munication à la même université
Seeyar, and B. Eidson. Frequency domain equal- en septembre 2010. Actuellement il
travail à l’école nationale des sciences
ization for single-carrier broadband wireless sys-
appliquées de khouribga au Maroc.
tems. IEEE Comm. Mag., 40(4):58–66, 2002.
Son domaine d’intérêt porte sur les techniques de codage
[5] J.L. Lacoume, P.O. Amblard, and P. Comon. conjoint source canal, les techniques MIMO pour les
Statistiques d’Ordre Supérieur pour le Traitement transmissions et les transmission vidéo H 264 dans le
contexte MIMO. Mail: fatani.imade@gmail.com
du Signal. Masson, 1997.
[6] R. K. Martin, J. Balakrishnan, W. A. Sethares,
and C. R. Johnson. A Blind Adaptive TEQ for
Multicarrier Systems. IEEE Signal Processing
Letters, 9(11):341–343, 2002.
[7] R. K. Martin and C. R. Johnson. Adaptive
Equalization: Transitioning from single-carrier to
multicarrier systems. in Proc. IEEE Signal mag.,
22(6):108–122, 2005.
[8] P. J. W. Melsa, R. C. Younce, and C.E. Rohrs.
Impulse Response Shortening for discrete Mul-
titone Transcievers. IEEE Trans. on comm.,
44(12):1662–1672, 1996.
[9] H. Minn, N. Al-Dhahir, and Y. Li. Opti-
mal Training Signals for MIMO OFDM Chan-
nel Estimation in the Presence of Frequency Off-
set and Phase Noise. IEEE Trans. on comm.,
54(10):1754–1759, 2006.
[10] T. Pollet, M. Peeters, M. Moonen, and L. Van-
dendorpe. Equalization for DMT-Based Broad-
band Modems. IEEE comm. mag., 38(5):106–
113, 2000.
[11] A.H. Sayed. Fundamentals of adaptive filtering.
John Wiley, 2003.
[12] B. Widrow, J. M. McCool, M. G. Larimore, and
C. R. Johnson. Stationary and Nonstationary
Learning Characteristics of the LMS adaptive fil-
ter. Proc. IEEE, 64(8):1151–1161, 1976.
51

ACSE Volume12 Issue2 P1111434342

Uploaded by

Copyright:

Available Formats

ACSE Volume12 Issue2 P1111434342

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ACSE Volume12 Issue2 P1111434342

Uploaded by

Copyright:

Available Formats

The International Congress for global Science

ICGST International Journal on Automatic

Volume (12), Issue (II)

© ICGST LLC, Delaware, USA, 2012

A Publication of the International Congress for

ICGST Editor in Chief: Dr. rer. nat. Ashraf Aboshosha

Improved PID MIMO Control of Turbo Generator relying on Genetic

Tarun Varshney, Satya Sheel

Abstract the nonlinear SISO Hammerstein model. Maja and Kabra

5. Problem Formulation Figure 2 Schematic diagram of PID tuning based on GA

different techniques is also shown. It can be seen that

7. Simulation Results Table2 Computational burden

Initially many individual solutions are randomly

where a, b are upper and lower limits of the parameter

is not possible in realistic time frame i.e. online /

Closed Loop Controller for Stochastic Systems with Uncertain Parameters

Mohamed F. Hassan and Hala A. Mourad

Abstract state estimation of linear discrete systems with uncertain

(8) yk 1  Cxˆ k 1|k  Cx k 1|k  C k 1 xˆ k 1|k  C k 1 x k 1|k  vk 1

k 11k 12k 21k 22k  

with   I  B k N 1B kT Prk 1

(RMSE) of the estimated states resulting from the

for k  1,..., k f and j  1,...., n (41)

where MI is the number of Monte Carlo iterations. 2

Moreover, the root mean square error index (RMSE 1

index) for each state is calculated using the following 2

RMSE index (x j )  (42) 0

Case xi Q R Pn% RMSE index 0

a 0.1 0.1 10 0.2597 0.269 20

b 0.5 1.0 20 0.7947 1.821

c 0.5 1.0 30 0.8235 8.233 4

d 1.0 1.0 30 0.8717 26.444 3

a 0.1 0.1 10 1.8834 1.887 1

b 0.5 1.0 20 95.214 125.57

c 0.5 1.0 30 299.092 345.97 0.7

From Table 1, the proposed filter has the best 5

performance in all cases. It has the minimum RMSE 0

whereas the (EKF) fails to estimate the states since it

diverges. The RMSE plots of the states using the three 1

filters for Case-a are shown in Figures (1-3), and for

Figure 1: RMSE of x1 (Case-a). Figure 5: RMSE of x 2 (Case-b).

Modified Kalman Filter k 1  1 0.005 0  k   0 

The system parameters KT , Kb , La and R ' , are

For the above model, it is assumed that the rotating

speed of the motor  is measured at the output. 8

6.2.a) State estimation under no control: 6

The continuous state equation of the DC motor is: 4

  Extended Kalman Filter

The above model is discretized using Euler’s method 5

Modified Kalman Filter

Extended Kalman Filter

Figure 11: Actual and Estimated motor angle

Figure 8: RMSE of (Case-b). displacement .

Figure 9: RMSE of (Case-b).

Figure 10: Block diagram of the closed loop system.

6.2.b) Observer-based Controller:

Figure 14: Feedback Control u.

7. Conclusion and System Engineering ACSE, vol. 7, no.2, pp1-

O. Chandra Sekhar, K.Chandra Sekhar

Abstract Tas, Tbs, T*cs The time equivalents of the voltage