Vous êtes sur la page 1sur 80

The 5th International Conference on the

Advancement of Science and Technology


ICAST-2017
Proceedings
Proceedings

Faculty of Electrical and Computer Engineering

Bahir Dar Institute of Technology


Bahir Dar University
May 2017

Bahir Dar, Ethiopia


Editors
Mr.Fanuel Melak……………………….... Chief Editor

Dr. Dr. Belachew Bantyirga………...…… Editor


Dr. Pusparashavan Annamalai………..… Editor

Dr. V. Balaji Visvanathan………………… Editor

Mr.Menargen Asmamaw……………….… Editor


TABLE OF CONTENTS

1 Optimizing Wind Power Extraction Using


ANFIS: Case Study with Ashegoda Wind
Farm ……………………………………. 1
Renald Chelladurai1, Leul Gebreselassie Asfaha,
Sujesh Kumar Kanaka Raj

2 Power Flow Analysis Using A Unified Power


Flow Controller ……………………………………. 13
Anil kumar Bonela, Assaye Bekele

3 On Board Charging Interface for Vehicle to


Grid (V2G) in Smart Grid Environment ……………………………………. 19
Dr.Santoshkumar, Kena Likassa Nefabas , Hinsermu
Alhemayu

4 Phase shifted carrier PWM based multilevel


converters for 1.5 MW wind turbines ……………………………………. 25
P.Palanivel, Tafese Asrat, Kena Likassa, Hinsermu
Alemayehu, Satish sukhavasi

5 Performance Analysis of Multiple Antenna


Based Blind Spectrum Sensing Techniques for
Cognitive Radio Networks ……………………………………. 31
Amare Kassaw1, Dr.Ing.Dereje Hailemariam2,
Prof. Dr.-Ing. Abdelhak M. Zoubir3

6 Joint Beamforming and Spectral Enhancement


for Robust ASR in Reverberant Environments
Fanuel Melak Asmare, Feifei Xiong, Mathias Bode, ……………………………………. 39
Bernd T. Meyer and Stefan Goetze

7 Modeling and Analysis of Linear Observer


Based On DC Machine ……………………………………. 47
Debabrata Pal1, Ejigu Tolina

8 Distributed Adaptive Cruise Control


……………………………………. 53
Solomon Genene Gudeta

9 Design of Controller via Feedback


Linearization for a Non-linear Conical Tank ……………………………………. 58
Dr.K.V.L.Narayana1, Kena Likassa Nefabas

10 Image processing for Pick and Place Gantry


Robot ……………………………………. 65
yonatan Tasew,daniel Nigusse
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

Optimizing Wind Power Extraction Using ANFIS: Case Study


with Ashegoda Wind Farm
Renald Chelladurai1, Leul Gebreselassie Asfaha2, Sujesh Kumar Kanaka Raj3
1
Dept. of Electrical & Comp. Engg., KIOT, Wollo University, Kombolcha, Ethiopia, renaldelectrical@gmail.com
2
Dept. of Electrical & Comp. Engg.,Mekelle University,Mekelle,Ethiopia, leul2000@gmail.com
3
Dept. of Electrical & Comp. Engg., KIOT, Wollo University, Kombolcha, Ethiopia, sujeshk1974@rediffmail.com

Abstract — Ethiopian Electric Power Corporation (EEPCO) was developing the first wind park
projects in the central and northern parts of Ethiopia. The amount of power output from a Wind
Energy Conversion System (WECS) depends upon the accuracy with which the peak power points are
tracked by the Maximum Power Point Tracking (MPPT) controller of the WECS control system
irrespective of the type of generator used. The influence of control modes on wind turbine efficiency in
maximum wind energy extraction has not received much attention so far. This paper focuses on the
influence of control modes on wind turbine efficiency in maximum wind energy extraction control
based on variable speed wind turbines with Doubly Fed Induction generator (DFIG) and Adaptive
Neuro-Fuzzy Inference System (ANFIS). ANFIS controller is applied in pitch control to extract
maximum power and rotor side converter for active power control and voltage regulation of a wind
turbine of DFIG. Wind turbine and its control unit are described in details. All power system
components and the ANFIS controller are simulated in Mathlab Simulink software. The Data used for
the comparison was taken from Ashegoda phase one wind farm project. The simulation results show
that the power extraction improves the efficiency by 72.82% when DFIG and ANFIS controller were
used and the grid stability has been also improved substantially.
Key words- DFIG, ANFIS, Slip, Controller and Energy loss
energy extraction control based on variable speed
I. INTRODUCTION wind turbines with Doubly Fed Induction Generator
(DFIG) (John Wiley, et al) and Adaptive Neuro-
As a result of shortage in conventional energy Fuzzy Inference System (ANFIS) controller. ANFIS
sources and increasing environmental concern, great controller is applied in pitch control to extract
efforts have been made to produce electricity from maximum power and rotor side converter for active
renewable sources, such as wind energy. Ethiopia power control and voltage regulation of a wind
has started to utilize its immense wind potential and turbine of DFIG (Meharrar et al, 2011 and
has recently connected many wind farms Mesemanolis et al, 2013). A wind turbine and its
(Ashegoda, Adama and etc.) with a total capacity of control unit are described in details. All power
171 MW. Ethiopian Electric Power Corporation system components and the ANFIS controller are
(EEPCO) was developing the first wind park simulated in Mathlab Simulink software (R2011).
projects in the central and northern parts of Ethiopia
with a total capacity of 120 MW. Simple, linear, The data used for the comparison was taken from
proportional-integral (PI) controllers are used in Ashegoda phase one wind farm project. The
these wind mills. The amount of power output from simulation results show that the power extraction
a Wind Energy Conversion System (WECS) improves the efficiency by 72.82% when DFIG and
depends upon the accuracy with which the peak ANFIS controller were used and the grid stability
power points are tracked by the Maximum Power has been also improved substantially.
Point Tracking (MPPT) controller of the WECS
control system irrespective of the type of generator II. OVERVIEW OF WIND FARM INDUSTRY
used (Mesemanolis et al). But there are many
Wind farms must be controlled in order to generate
factors that influence the wind turbine efficiency, electricity when the wind speeds are sufficient and
such as electrical characteristics of the generator, disconnected for low wind speed or very high wind
aerodynamic characteristics of the turbine blades speed conditions. In addition, the reactive power
and maximum power extraction control strategies demand of the wind farm and the voltage at the
(Abdul Ghani et al). However, the influence of point of common coupling must also be controlled
control modes on wind turbine efficiency in under all operating conditions such as normal, fault
maximum wind energy extraction has not received conditions, over-loading operation, and islanding.
much attention so far (Perdana et al, 2004). Generally, wind farms are equipped with wind
turbine control systems and a central automatic
This paper focuses on the influence of control
control system. A wind turbine control system,
modes on wind turbine efficiency in maximum wind
1
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

consisting of the unit control and the local control,


receives information about the wind (speed and (1)
direction), turbine temperature, vibration level,
generator voltage, current and rotational speed and B. Mechanical System
the utility interconnection status. According to the The mechanical elements of a wind turbine and the
received information, the turbine control system forces suffered or transmitted through its
performs several control operations such as turbine components are very numerous. It is therefore
operation start or stop control, brake control for necessary to choose the dynamics to represent and
emergencies, pitch angle control for variable pitch the typical values of their characteristic parameters.
turbines, yaw drive control for upwind turbines,
The first is the resonant frequency of the power
generator synchronizing control for synchronous
train. The power transmission train is constituted by
generators (where there is no interface), voltage
regulation, reactive power control, and turbine the blades linked to the hub, coupled to the slow
transformer on-load tap changer control. A central shaft, which is linked to the gearbox, which
control and monitoring system is required for wind multiplies the rotational speed of the fast shaft
farms, especially large farms with large arrays of connected to the generator. A two mass model is
wind turbines. Usually, this is achieved by using a illustrated in Fig. 2 can then model the drive train.
Supervisory Control and Data Acquisition
(SCADA) system. It allows a central computer The second, the resonance frequency is much higher
system to monitor and control each turbine and its magnitude is lower. All the magnitudes are
operation as well as the entire wind farm operation. considered in the fast shaft. Inertia concerns the
The SCADA is located in the central room, the turbine side masses, while Jm concerns those of the
substation of the wind farm or even at a remote off electrical machine. These inertias do not always
site point. represent exactly the turbine and the electrical
machine. If the fundamental resonance frequency
III. MODELING AND CONTROL OF VARIABLE SPEED comes from the blades, part of the turbine inertia is
WIND TURBINE then considered in Jm.
Variable Speed Wind Turbine (VSWT) modeling is The stiffness and damping coefficients, Ktm and Dtm,
composed of Aerodynamic model, Pitch system, define the flexible coupling between the two
Mechanical system, Electrical machine, Power inertias. As for the inertias, these coefficients are
converters and Control system (John Wiley, et al). not always directly linked to the fast shaft, but to the
wm
fundamental resonance, which may be located
Tt Mechanical
Wind
Model
Vv Aerodynamic
System
System
Tem
Electrical
Machine somewhere else. Dt and Dm are the friction
wr Grid

Vgenerator
Vgrid
Model
coefficients and they represent the mechanical
β Converters
losses by friction in the rotational movement (John
Pitch Control
System
Vconverter Wiley, et al).
Torque &
Reactive

βref
Power Control
The turbine rotational speed and driving torque are
Tem_ref
Qref
expressed in the fast shaft by (2) and (3)
Wind
Turbine Control
Vv Vgrid
Strategy
(2)
Fig.1 Block scheme of a variable speed wind turbine
model (3)
A. Aerodynamic Model Where N is the gearbox ratio
The aerodynamic model represents the power
extraction of the rotor, calculating the mechanical
torque as a function of the air flow on the blades.
The wind speed can be considered as the averaged
incident wind speed on the swept area of the blades
with the aim of evaluating the average torque in the
lower speed axle.
The torque generated by the rotor has been defined
by (1)

2
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

Wind energy transfer


2. Follow the curve of maximum power
wt_ar wm
extraction from variable speed operation
Ktm
with partial load.
Tt_ar Tem 3. Limit the maximum speed at partial load
Jt Jm
operation.
Dtm 4. Limit the maximum operating speed at rated
Dt Dm
power output.
The minimum speed limit is explained by the fact
that we must prevent the turbine from rotating at
Fig.2 Two mass mechanical model. speeds corresponding to the resonant frequency of
C. Control of a Variable Speed Wind Turbine the tower. This resonance frequency is about 0.5 Hz
and a small rotational speed can excite it. Moreover,
Control of a VSWT is needed to calculate the for DFIM based turbines this limitation also serves
generator torque and pitch angle references in order to limit the sliding of the electrical machine, and
to fulfill several requirements like extract the hence the rotor voltage, and therefore the voltage
maximum energy from the wind, keep the turbine in that must provide the drive rotor.
safe operating mode and minimize mechanical loads
in the drive train. Therefore, the wind generator starts to run at the
wind speed connection (cut-in wind speed) with a
Fig.3 shows a general control scheme for the rotating speed Ωt-min.
VSWT, where the two degrees of freedom are the
generator torque and the pitch angle. When the wind speed becomes more important, it
reaches the maximum aerodynamic performance
operating in Zone 2. As wind speed increases, the
rotation speed also increases until the maximum
rotation speed Ωt-max. The wind generator then
operates in Zone 3. When wind speed reaches its
nominal value, the generator works at the rated
mechanical power and the energy captured for
higher wind speeds should be regulated at this
nominal value.

Fig. 3. General control scheme for the VSWT


This control is independent of the generator
technology and can be simulated without modeling
the electrical machine, power converters, and their
associated controls just including the torque
dynamics as a first-order system. Moreover, for
DFIG based wind turbines this limitation also serves
to limit the slip of the electrical machine and
therefore the A.C. voltage must be provided to the
rotor converter.

D. Turbine Speed Control Regions Fig.4. Wind turbine control strategy


The wind turbine control strategy most commonly Zone 4 corresponds to operation at full load. Here,
used is illustrated in Fig.4 and consists of four the mechanical power can be limited either by
operation zones as follows varying the pitch or by torque control. Typically, the
1. Limit the minimum speed of operation. electromagnetic torque is maintained at nominal

3
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

value and adjusts the pitch angle to keep the turbine constant DC-link voltage. The GSC‘s reactive
at maximum speed and rated power. power generation is not used as the RSC
independently does. But, during the steady state and
E. Pitch System Controller
low voltage periods, the GSC is controlled to take
Pitch System controller is designed for rotating all part in the reactive power generation. The GSC
the blades at the same angle or each of them supplies the reactive current quickly while the RSC
independently. This independent regulation gives results in delays as it passes the current through the
more degrees of freedom to the control system. This machine. These converters can temporarily be
particular operation would reduce the stresses in the overloaded, so that during short circuit periods, the
blades. The independent regulation of blades is an DFIG can make a better contribution to the grid
important innovation that will bring more voltage.
intelligence into the control system of wind
Power flow of the rotor is bidirectional. When
turbines.
, the power flows from the rotor to the
In studying a dynamic control system, a blade pitch power grid and when , the rotor absorbs the
involves many torques and forces. The energy from the power grid. Power electronic
representation of these torques requires modeling converters between the rotor and adjust the grid
the structural dynamics of the blade, the behavior of frequency and amplitude of the rotor voltage. The
the air around the blades, or the inclusion of friction control of the rotor voltage allows the system to
in the bearings. Moreover, regulation of the speed of operate at a variable-speed while still producing
rotation around the longitudinal axis of the blades constant frequency electricity. The mechanical
has a bandwidth much greater than that of the power and the stator electric power output are
control of the angle itself. The most standard computed as in (4.a, b).
approach is to represent the loop control, the rate of
change of pitch angle, and a linear system of first (4.a)
order containing the main dynamics of the actuator
(hydraulic or electric). In fact, when modeling the (4.b)
pitch control, it is very important to model the rate For a lossless generator the mechanical equation is
of change of this angle. Indeed, given the effort
sustained by the blades, the variation of the pitch (5)
must be limited. It is limited to about 100/s during
normal operation and 200/s for emergencies. In steady-state at fixed speed for a lossless generator
is
Regulation of the pitch system is modeled as shown
in Fig.5, by a PI controller that generates a reference (6)
rate of change of pitch. This reference is limited and
a first order system gives the dynamic behavior of (7)
speed control of pitch variation. The pitch angle
itself is then obtained by integrating the variation of
the angle (John Wiley, et al) .
(8)
β* PI dβ* 1 dβ 1 β
+
Controller Tdβs+1 s The slip of the generator.
-
(9)
Where, Pm is the extracted mechanical power.
Fig.5. Pitch system and control model. Ps is the power from the stator to the grid.
F. Power Flow Pr is the power from the rotor to the grid.
ωr is the rotor rotational speed.
The grid connected DFIG is the most reliable ωs is the synchronous speed.
system to harness the wind power. As the DFIG J is the combined rotor and wind turbine
utilizes the turns ratio of the machine, the converter inertia coefficient.
need not to be rated for the machine‘s full rated
power. The Rotor Side Converter (RSC) controls i. Rotor side converter controller
the active and reactive power of the machine while
the Grid-Side Converter (GSC) maintains the
4
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

For the RSC controller the d-axis of the rotating Fig.6. Rotor side converter controller
reference frame used for d-q transformation is
ii. Grid side converter controller
aligned with air-gap flux. The actual electrical
output power, measured at the grid terminals of the Vdc_ref
wind turbine, is added to the total power losses Vdc DC Voltage Idgc_ref
(mechanical and electrical) and is compared with -
Regulator
+

the reference power obtained from the tracking


characteristic. A Proportional-Integral (PI) regulator -
Current Vgc
is used to reduce the power error to zero. The output Idgc Regulator
Igc
of this regulator is the reference rotor current Iqr_ref Current
Iqgc
Measurement
that must be injected in the rotor by converter Crotor. -
+
This is the current component that produces the Iq_ref

electromagnetic torque Tem. The actual Iqr


Fig.7. Grid side converter controller
component is compared to Iqr_ref and the error is
reduced to zero by a current regulator (PI). The The GSD is used to regulate the voltage of the DC
output of this current controller is the voltage Vqr bus capacitor. For the GSD controller the d-axis of
generated by Crotor. The current regulator is assisted the rotating reference frame used for d-q
by feed forward terms which predict Vqr. The transformation is aligned with the positive sequence
voltage at grid terminals is controlled by the of grid voltage. This controller consists of
reactive power generated or absorbed by the
converter Crotor. The reactive power is exchanged 1. A measurement system measuring the d and q
between Crotor and the grid, through the generator. In components of AC currents to be controlled as well
the exchange process the generator absorbs reactive as the DC voltage Vdc.
power to supply its mutual and leakage inductances. 2. An outer regulation loop consisting of a DC
The excess of reactive power is sent to the grid or to voltage regulator.
Crotor.
3. An inner current regulation loop consisting of a
When the wind turbine operates in a VAR current regulator. The current regulatory controls
regulation mode, the reactive power at grid the magnitude and phase of the voltage generated by
terminals is kept constant by a VAR regulator. The converter Cgrid (Vgc) from the Idgc_ref produced by the
output of the voltage regulator or the VAR regulator DC voltage regulator and specified Iq_ref reference.
is the reference d-axis current Idr_ref that must be The current regulator is assisted by feed forward
injected in the rotor by converter Crotor. The same terms which predict the Cgrid output voltage (John
current regulator as for the power control is used to Wiley, et al) .
regulate the actual Idr component of a positive -
sequence current to its reference value. The output G. Adaptive neuron-fuzzy inference system
of this regulator is the d-axis voltage Vdr generated In recent years, fuzzy logic control has played an
by Crotor. The current regulator is assisted by feed increasing and significant role in the development
forward terms which predict Vdr. Vdr and Vqr are and design of real-time control applications.
respectively the d-axis and q-axis of the voltage Vr However, membership function type, number of
(John Wiley, et al) . rules and correct selection of parameters of fuzzy
Vref
controller are very important to obtain desired
V Me a
AC
sure
Vac
+
AC performance in the system. Determination of
me n Regulator
t
- membership function type and rule number of fuzzy
I Power
Measurement Vxs
Qref
Idr_ref
+
controller and selection of parameters is made by
+
V
I
Var
Measurement
Q
-
Var
Regulator - means of trial and error method and by using the
Current Vr specialization knowledge.
Regular
Idr
Ir Current
Measurement
Iqr
Adaptive Neuron-Fuzzy Inference System is the
wr Tracking
Characterstics Pref
- integration of artificial neural networks and fuzzy
+
V
Power P
+
Power Iqr_ref
inference systems. ANFIS is formulated on three
I Measurement - Regulator

wr -
main elements that auxiliary, compatible and
Power
Loss
Pl integrative. ANFIS is also expressed as functional
Igc
adaptive networks unit equivalent to fuzzy inference
system. ANFIS is the combination of neural
5
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

networks and fuzzy system to determine parameters


of the fuzzy system. The main purpose of using the
Neuron-Fuzzy approach is to automatically realize
the fuzzy system by using the neural network
methods.
In ANFIS control system, Fuzzy Segno models are
involved in framework of adaptive system to
facilitate the learning and adaptation studies. ANFIS
permits combination of numerical and linguistic
data. Besides, Neuron-Fuzzy systems have the
ability to obtain fuzzy information from numerical
data.
In the adaptive neuron-fuzzy model, two basic
learning algorithms are required. One is the
structural learning algorithm to find suitable fuzzy Fig.8. General Schematic diagram controller system
logic rules and the other is the parameter learning
Rule 1: If (x is A1) and (y is B1) then f1 = p1x+q1y+r1
algorithm to adjust the membership functions and
other parameters according to desired performance Rule 2: If (x is A2) and (y is B2) then f2 = p2x+q2y+r2
from the system. ANFIS is a multi-layer adaptive
Where x and y are the inputs, Ai and Bi are the
neural network based fuzzy inference system.
fuzzy sets, fi are the outputs within the fuzzy region
ANFIS algorithm is composed of fuzzy logic and
specified by the rule, pi, qi and ri are the design
neural networks with 5 layers to implement different
parameters that are computed during the training
node functions to learn and tune parameters in a
process. Among the five layers, the first and fourth
Fuzzy Inference System (FIS) structure using a
layers consists of adaptive nodes while the second,
hybrid learning mode. In the forward pass of
third and fifth layers consist of fixed nodes. The
learning, with fixed premise parameters, the least
adaptive nodes are associated with their respective
squared error estimate approach is employed to
parameters, get updated with each subsequent
update the consequent parameters and to pass the
iteration while the fixed nodes are devoid of any
errors to the backward pass. In the backward pass of
parameters.
learning, the consequent parameters are fixed and
the gradient descent method is applied to update the Layer 1: Every node in the layer 1 is an adaptive
premise parameters. The premise and consequent node. The outputs of layer 1 are the fuzzy
parameters will be identified for Membership membership grade of the inputs, which are given
Function (MF) and FIS by repeating the forward by:
and backward passes. ANFIS is a fuzzy Sugeno
model put in the framework of adaptive systems to (10)
facilitate learning and adapting. Such framework
makes the fuzzy logic controller more systematic (11)
and less relying on expert knowledge (Meharrar et Where x and y are the inputs, where A, B is a
al, 2011 and Mesemanolis et al, 2013). linguistic label (large, small) and , can
To show the ANFIS architecture, let us take two- adopt any fuzzy membership function.
fuzzy rules based on a first order Segno model.
Layer 2: It is a rule layer, calculates the strength for
each rule quantifying the extent which any input
data belongs to that rule.
It is a fixed node labeled MW hose output is the
product of all the incoming signals, the outputs of
this layer can be represented as,

(12)

6
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

Layer 3: It is normalization layer in this also fixed IV. STUDY OF ASHEGODA WIND FARM
node is a circle node labeled N. Ashegoda wind farm site is situated in the northern
Every node in this layer calculates the ratio of the ith Ethiopian highland, about 775 km from Adiss
rule’s firing strength to the sum of all rule’s firing Abeba and 15 km from Mekelle at 13° 25' 31.23"N
strengths. 39° 34' 21.20"E (Latitude / Longitude) and at an
altitude of 2400 m above sea level close to the
(13) descent to the coastal plain. The whole area,
foreseen for the construction of the wind farm, is
Layer 4: It is De-fuzzification layer an adaptive covered with small bushes and grass. The land is
node with a node, the output of each node in this mainly used for extensive goat farming, and partly
layer is simply the product of the normalized firing for agricultural use. The geographical terrain
strength and a first order polynomial. conditions can be classified as medium complex at
the highland and complex for descend to the coastal
plain.
The total average power production for 2013 is
(14) 45803MWh and the total yearly calculated power
Layer 5: It is summation neuron a fixed node which production is 89693MWh and the total year energy
computes the overall output as the summation of all loss is 43890MWh as listed in TABLE I.
incoming signals. Table I. 2013 YEARLY POWER PRODUCTIONS VS.
MONTHLY WIND SPEED
(15)
Avg. Energy Cal. wind Cal. Energy
wind productio power (kw) Energy loss
There are two methods that ANFIS learning Mon
power n(MW) productio (MWh)
employs for updating membership function (KW) n(MWh)
th
parameters: Jan. 5454.136 3916 9564.295 7115.835 3199.835
Feb. 7288.306 4647 15026.47 10097.79 5450.79
1. Back propagation for all parameters (a Mar. 6114.866 4012 15886.31 11819.42 7807.418
steepest descent method). Apr. 7287.709 4943 12084.71 8700.989 3757.989
2. A hybrid method consisting of back may 4812.807 3447 7399.643 5505.334 2058.334
Jun. 3235.454 2263 5480.95 3946.284 1683.284
propagation for the parameters associated
Jul. 7079.414 5108 12036.42 8955.095 3847.095
with the input membership functions and Aug
least squares estimation for the parameters . 1933.481 1403.75 3810.931 2835.333 1431.583
associated with the output membership Sep. 2051.519 1439.32 3856.49 2776.673 1337.353
Oct. 5315.232 3857.63 9767.885 7267.307 3409.677
functions. Nov
. 7560.536 5381.93 14269.1 10273.75 4891.824
In order to improve the training efficiency, a hybrid
Dec. 7349.967 5384.674 13976.93 10398.83 5014.16
learning algorithm is applied to justify the Total 65483.43 45803.3 123160.1 89692.64 43889.34
parameters of input and output membership
functions. In this way a two-step process is used for In November month, from 30 wind turbines the
the learning or adjustment of the network maximum power generated is 300KW at wind speed
parameters. In the first step, the premise parameters of 7.66m/s.
are kept fixed and the information is propagated The total power production for this month is
forward in the network to Layer 4, where the 5,382MWh and the total calculated power
consequent parameters are identified by a least- production is 10,274MWh and the total energy loss
squares estimator. In the second step, the backward for this month is 10274-5382=4,892MWh as listed
pass, the consequent parameters are held fixed while in TABLE II.
the error is propagated and the premise parameters
are modified using a gradient descent algorithm. Table II. EEPCO Ashegoda Production Data For The
The only user-specified information is the number Month Of November
of membership functions for each input and the Turbin
e No.
Avg.
wind
Avg.
wind
Energy
production
Cal. wind
power
Cal.
Energy
Energy
loss
input–output training information (Meharrar et al, speed power (MWh) (kw) production (MWh)
(m/s) (KW) (MWh)
2011) . 1 8.04 251.781 180.7769 619.6587 446.1543 265.3774
2 7.51 272.7587 189.1838 505.0149 363.6108 174.4269
3 7.51 282.1987 202.8979 505.0149 363.6108 160.7128
4 7.66 295.9555 210.2051 535.8839 385.8364 175.6313
5 7.62 283.7299 203.8989 527.5326 379.8235 175.9246

7
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

6 7.52 276.5959 195.0898 507.035 365.0652 169.9754 690/33kV 25kV/230kV


7 7.87 281.5933 196.8911 581.1773 418.4476 221.5565 12MW 65MW
8 7.47 281.716 199.6951 496.9884 357.8316 158.1366
9 7.65 261.2919 187.754 533.7879 384.3273 196.5733 DFIG
10 7.4 269.2465 193.6892 483.1473 347.8661 154.1769 15km
Δ/Υ
11 7.47 276.1377 198.6929 496.98837 357.83163 159.13876 B690 Δ/Υ B33 B230
(690v) (33k) (230kV)
12 7.85 287.6149 206.001 576.7577 415.2655 209.2646
Load
13 7.27 268.5268 192.2869 458.1289 329.8528 137.5659 500kw
14 6.82 227.3612 163.259 378.214 272.314 109.055
15 7.03 231.9752 165.3611 414.2384 298.2517 132.8906
16 7.13 235.1658 169.1609 432.1684 311.1613 142.0004
17 7.1 231.6086 166.6619 426.7362 307.2501 140.5882
18 7.03 218.6016 157.2625 414.2384 298.2517 140.9892
19 7.35 273.191 193.1747 473.4198 340.8623 147.6876
20 7.44 200.1877 135.2312 491.0246 353.5377 218.3065
21 7.21 245.7285 169.2646 446.8793 321.7531 152.4884
22 7.39 193.9585 136.2668 481.1913 346.4577 210.1909
23 6.79 208.8089 149.8455 373.2448 268.7362 118.8908
24 7.07 227.6847 163.5591 421.3497 303.3718 139.8127
25 7.18 246.4637 170.4567 441.3242 317.7534 147.2968
26 7.37 260.4142 187.3828 477.295 343.6524 156.2696
27 7.1 242.1925 174.2705 426.7362 307.2501 132.9795
28 7.06 231.5331 166.563 419.5643 302.0863 135.5233
29 7.39 273.6049 196.7913 481.1913 346.4577 149.6665
30 7.19 222.9098 160.3562 443.1708 319.0829 158.7267
Total 220.49 7560.536 5381.93 14269.1 10273.75 4891.824

V. SIMULATION MODEL
A 6 MW wind farm consisting of six 1 MW wind
turbines connected to a 33 kV distribution system
exports power to a 230 kV grid through a 15 km, 33
kV feeder is considered. A 500 KW resistive load is
connected at the 690 V generation bus. Wind
turbines using a doubly-fed induction generator
consist of a wound rotor induction generator and an
AC/DC/AC IGBT-based PWM converter. The
Fig.9. Simulation model of wind farm connected to a
switching frequency is chosen to be 2 kHz. The
distribution system
stator winding is connected directly to the 50 Hz
grid while the rotor is fed at variable frequency
through the AC/DC/AC converter. The DFIG A. ANFIS Controller Design
technology allows extracting maximum energy from
the wind for low wind speeds by optimizing the After wind power generation system is
turbine speed, while minimizing mechanical stresses commissioned, electrical output magnitudes become
on the turbine during gusts of wind. The optimum very important with regard to system productivity
turbine speed producing maximum mechanical and reliability. If the power obtained from the wind
energy for a given wind speed is proportional to the power generation system is desired to be high
wind speed. quality, output voltage and frequency must be
within operation limit values. For this purpose, to
obtain electrical power in desired quality from the
power generation system, operation voltage,
frequency and power has been controlled with
ANFIS.
In the modeling and feedback control of any
dynamical system, a controller is a must for the
plant as it takes care of all the disturbances and
brings back the system to its original state in a
couple of second.
To start with, we design the controller using ANFIS
scheme. The model considered here is based on
Takagi- Sugeno fuzzy inference model. The block
diagram of the proposed ANFIS control scheme is
shown in Fig.10.

8
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

+ ANFIS WIND
CONTROLLER TURBINE
Reference
-
Measured

Fig.10. Block diagram of the ANFIS control scheme


The inputs to the ANFIS Controller are error and
change in error .
Fig.11 shows the FIS editor of Sugeno type Fuzzy
Inference System with two inputs (error and change
in error) and one output. Each input is having nine
linguistic variables with triangular membership
functions. 81 rules are framed. The rules are viewed
by rule viewer as shown in Fig.12

Fig.13. ANFIS structure for the control of wind turbine


B. Rotor side converter controllers
Fig.14 shows DFIG control in wind energy system.
The figure describes the block diagram of rotor side
converter to which adaptive neuron-fuzzy (ANFIS)
controller is applied. The main objectives of this
part are active power control and voltage regulation
of DFIG wind turbine using output reactive power
control. As illustrated in Fig.14 rotor side converter
manages to follow reference active (Pref) power and
voltage (Vref) separately using ANFIS controllers.
Fig.11. FIS editor (Sugeno Model) with two inputs and After the production of reference d and q-axis rotor
one output. currents, they converted to a-b-c reference frame
using flux angle, rotor angle and finally slip angle
calculation and Concordia and Park transformation
matrix. Then they applied to a hysteresis current
controller to be compared with actual currents and
produce switching time intervals of converter.

Fig.12. Rule Viewer of Fuzzy Sugeno Model


After running this FIS file with simulink model, the
training data are collected and loaded by the ANFIS
editor. The ANFIS structure is trained with hybrid
learning up to 50 epochs, with error tolerance of
zero. ANFIS structure for the control of wind
turbine is shown in Fig.13

9
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

angle starts to increase to limit the speed, as in


Fig.16

Fig.15. Waveforms of DFIG wind energy during fault at


Bus B33 (Voltage Regulation Mode).

(a)

Fig.16. Waveforms of DFIG wind energy during fault at


(b) Bus B33 (Var Regulation Mode).
Fig.14. DFIG control in wind energy conversion system
(a) Voltage regulating side (b) Rotor side
In this simulation, we observe the impact of a single
phase-to-ground fault occurring on the 33 kV line.
At t=5 s a 9 cycle (0.15 s) phase-to-ground fault is
applied on phase A at B33 bus. When the wind
turbine is in voltage regulation mode, the positive
sequence voltage at wind turbine terminals
(V1_B690) drops to 0.8 p.u. during the fault, which
is above the under voltage protection threshold
(0.75 p.u. for a t>0.1 s). The wind farm therefore
stays in service, as shown in Fig.15. However, if the
VAR regulation mode is used with Qref=0, the
voltage drops under 0.7 p.u. and the under voltage
protection trips the wind farm. We can now observe
that the turbine speed increases. At t=40s the pitch

10
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

We select one month data which is november from


Ashegoda wind farm among the twelve data. Then
we compared that with the simulated result that
shown in Fig.18 and Fig.19.
Table III. Ashegoda Power Production Data And
ANFIS Controlled DFIG PowerProduction Data For The
Month Of November
wind Power ANFIS
Variation Efficiency
speed(m/s) Measured Controller
6.79 208.8089039 927 718.1911 77.47476765
6.82 227.3611661 927 699.63883 75.47344486
7.03 231.9751789 927 695.02482 74.97570885
7.03 218.6016425 927 708.39836 76.41837729
7.06 231.5330768 927 695.46692 75.02340056
7.07 227.6846807 927 699.31532 75.43854578
7.1 231.6086254 927 695.39137 75.01525076
7.1 242.1924935 927 684.80751 73.87351743
7.13 235.1657685 927 691.83423 74.63152444
7.18 246.4637321 927 680.53627 73.41275813
7.19 222.9098429 927 704.09016 75.95363075
7.21 245.7284927 927 681.27151 73.49207199
7.27 268.5268359 927 658.47316 71.03270379
7.35 273.1910285 927 653.80897 70.52955464
Fig.17. ANFIS speed regulator and pitch control model. 7.37 260.4141943 927 666.58581 71.9078539
7.39 193.9584924 927 733.04151 79.07675379
7.39 273.6049455 927 653.39505 70.4849034
In Fig.17 ANFIS control scheme for extracting 7.4 269.2465219 927.1666667 657.92014 70.96028885
maximum power from a variable speed wind turbine 7.44 200.1876533 927.1666667 726.97901 78.40866583
has been presented. It has been shown that the 7.47 281.7159783 927.1666667 645.45069 69.61538972
7.47 276.1372676 927.1666667 651.0294 70.2170842
turbine power output depends nonlinearly on its 7.51 272.7587023 927.3333333 654.57463 70.58676827
angular speed and the wind speed. ANFIS control is 7.51 282.1987203 927.3333333 645.13461 69.56879365
7.52 276.5959165 927.3333333 650.73742 70.17297809
well suited for searching the optimum speed at 7.62 283.7299049 927.3333333 643.60343 69.40367668
which the turbine should operate under varying 7.65 261.2919408 927.3333333 666.04139 71.82329898
7.66 295.9555417 927.3333333 631.37779 68.08531183
wind conditions. In order to verify that maximum 7.85 287.6149269 927.3333333 639.71841 68.9847311
power is extracted from the available wind, power 7.87 281.5932854 927.3333333 645.74005 69.63408138
8.04 251.7809693 927.3333333 675.55236 72.84892495
coefficient Cp (λ) has to be observed. 7.349666667 252.017881 927.1222222 675.10434 72.81749205

Fig.18. Simulation result of the ANFIS controller Fig.19. Wind power production improvement

11
Optimizing Wind Power Extraction Using ANFIS: Case Study with Ashegoda Wind Farm
ICAST-2017

VI. CONCLUSION www.elsevier.com/locate/eswa, Expert Systems with


Applications, vol.38,pp. 7659–7664.
In this paper optimum doubly-fed induction
generator (DFIG) ANFIS control of wind turbine in 3. A. Mesemanolis and C. Mademlis, Self-Tuning
order to extract maximum power is described and Maximum Power Point Tracking Control for Wind
verified through the simulation. Generation Systems.
4. A. Mesemanolis, and C. Mademlis,“ On-line
The wind turbine driven by doubly-fed induction estimation of induction generator parameters using
machine is a part of distributed generation which adaptive neuro-fuzzy inference systems for wind
feeds AC power to the distribution network. The
energy conversion systems,” International
system is modeled and simulated in the Matlab
Conference on Renewable Energies and Power
Simulink environment in such a way that it can be
Quality(ICREPQ‟13) Bilbao (Spain),20th to 22th
suited for modeling of all types of induction
march,2013.
generator configurations.
5. A. Perdana, O. Carlson, and J. Persson, “Dynamic
The main goal of implementing ANFIS controller is Response of Grid-Connected Wind Turbine with
to continuously adapt the rotational speed of the Doubly Fed Induction Generator during
generator to the wind speed in a way that the turbine Disturbances,” Nordic workshop on power and
operates at its optimum level of aerodynamic industrial electronics, Trondheim – 2004.
efficiency. The advantages of using ANFIS 6. Abdul Ghani Abro And Junita Mohamad-Saleh,”
controller are universal control algorithm, fast control of power system stability reviewed solutions
response, and parameter insensitivity. For based on intelligent systems,” International journal
comparison, the data was taken from Ashegoda of innovative computing, information and control
phase one wind farm project. The simulation results volume 8, numb
show that the power extraction when DFIG and
ANFIS controller were used, improves the
efficiency by 72.82% and the grid stability has been
also improved substantially. So that this kind of cost
effective, optimized results helps Ethiopia, to
produce more wind power.
The completion of one research project opens the
avenues for work in many other related areas. The
following areas are identified for future work:
1. The same work can be extended to the second,
third Ashegoda wind farm and in the other
Ethiopia wind farm.
2. Power quality improvement by ANFIS GSC
controlling for filtering the harmonic currents
of a nonlinear load with a harmonic distortion
ratio.
3. Dynamic modeling and control ANFIS of wind
farms with more than one wind turbine and their
interaction with each other in case of unequal
wind distribution.

REFERENCES
1. A.John Wiley & Sons, “Doubly fed induction
machine: Modelling and control for wind energy
generation,” ISBN 978-0-470-76865-5.
2. A. Meharrar , M. Tioursi, M. Hatti, & A.
Boudghène Stambouli, 2011, “A variable speed wind
generator maximum power tracking based on
adaptive neuron-fuzzy inference system” available at

12
Power Flow Analysis Using A Unified Power Flow Controller
ICAST-2017

Power Flow Analysis Using A Unified Power Flow Controller


Anil kumar Bonela1, Assaye Bekele1
1
Department of Electrical and Computer Engineering, Dire Dawa university, Dire Dawa, Ethopia
anilk193@gmail.com, , assayeeo@yahoo.com

Abstract --- This paper presents the power flow control in electric power systems by use of an
improved steady state mathematical model of unified power flow controller embedded in a power
system. The main characteristic of the approach is that an equivalent mathematical model is
developed based on the concept of injected powers in which the operational losses can be taken into
account. The model is quite suitable in load flow studies, since it accepts employing conventional
techniques such as a Newton Raphson method. The model is validated by embedding it in IEEE 14 bus
system and then carrying out the load flow studies using MATLAB. The results of load flow analysis
show the effectiveness of the model.
Key words - FACTS, UPFC, power flow analysis
true representation of them in a computational
I. Introduction environment. Converters are modeled as
controllable voltage sources, while the effects of the
The power transmitted over an AC transmission line transformers are modeled as pure inductances
is a function of the line impedance, the magnitude connected to the lines and real power loses in
of sending-end and receiving-end voltages, and the UPFC.
phase angle between these two voltages. There is a
need for new power flow controllers capable of
increasing transmission capability and controlling II. Steady-State UPFC Representation
the parameters affecting the power flow in the The conceptual hardware configuration of UPFC is
transmission line (Gyugyi, 1992). Flexible AC shown in Fig. 1. Converters labeled as “series
transmission system (FACTS) devices give more converter” and “shunt converter” are operated from
flexibility of control for secure and economic a common DC link voltage provided by DC storage
operation of power systems. Among FACTS capacitor. Two coupling power transformers are
devices, the unified power flow controller (UPFC) also required to isolate UPFC and the transmission
is emerging as a promising solution for improving line, and to match the voltage levels between the
power system characteristics for its high degree of power network and voltage produced by the
controllability of many power system variables. converters. This arrangement can be functionally
UPFC can control simultaneously or selectively, all treated as an ideal AC to AC power converter in
parameters affecting power flow in the transmission which the magnitude and phase shift of the AC
line, i.e. voltage, impedance and phase angle. It can output voltages of both converters can be controlled
also independently control both real and reactive at any desired value, assuming that the controlled
power flow in the transmission line, besides that it voltage source in series with the transmission line
has the capabilities of improving transient stability, can be controlled without restriction. This means
mitigating system oscillations and providing voltage that the phase angle of the series injected voltage
support. Performance analysis of UPFC in load flow can be chosen independently of the line current.
studies requires its steady state modeling (Nabavi- Eventually as seen in Fig. 1, the real power can
Niaki, 1996; Ambriz-Perez & Fuerte-Esquivel, freely flow in either direction between ac terminals
2000). In Fuerte-Esquivel (1997), UPFC is of the two converters and each converter can also
represented by two ideal voltage sources with series generate or absorb reactive power independently at
source impedances, connected in series and parallel its own AC output terminals. The series converter
with the transmission line, representing the output performs the main functions of UPFC, while the
voltages of series and shunt branches of UPFC. shunt converter is used to provide real power
Because UPFC employs two voltage source demanded by the series converter and the losses in
converters and two coupling transformers, the UPFC.
mathematical model proposed here is based on the

13
Power Flow Analysis Using A Unified Power Flow Controller
ICAST-2017

Fig. 1. Schematic diagram of an UPFC

The UPFC can be represented in steady state by the Fig. 3. Phasor diagram
two voltage sources with appropriate impedances as The related phasor diagram of the concerned
shown in Fig. 2. The voltage sources can then be parameters in (1) and (2) is drawn in Fig. 3.
represented by the relationship between the voltages
and amplitude modulation ratios, and phase shifts of In Fig. 3, voltage of bus i, Vi is assumed to be
UPFC. In this model, the shunt transformer
reference vector i.e vi  vi  00 . The power
impedance and the transmission line impedance
including the series transformer impedance are injection model can be obtained by replacing the
assumed to be constant. The mathematical model is voltage source Vse by a current source Ise in parallel
constructed by representing the AC output terminals with the transmission line as shown in Fig. 4.
of the two converters with two ideal voltage sources,
I se   jbseVse (3)
Vse and Vsh respectively in series with the
reactance’s Xse, and Xsh, denoting the leakage Where bse  1/ X se
reactance of the two coupling transformers
respectively in Fig. 2. IL represents transmission line
current having a phase angle of IL

Fig. 4. Replacement of series voltage source by a current


source
The effects of the current source Ise and susceptance
bse can be modeled by the injection powers at buses i
and j.
Fig. 2. Equivalent circuit of UPFC Sis  Vi (I se )* (4)
A. Series Connected Voltage Source Converter S js  V j (  I ) * (5)

As seen in Fig. 2, Vi represents an imaginary Sis  Pis  jQis 


(6)
voltage behind the series reactance Xse rbse vi2 sin   jrbse vi2 cos 
Vi  Vse  Vi (1) S js  Pjs  jQ js 
Series voltage source, Vse is controllable both in vi v j bse r sin i   j    (7)
magnitude and phase angle
 jvi v j bse cos i   j   
Vse  rVi ei (2)
Based on (6) and (7), power injection model of
series connected voltage source can be seen as two
Where, 0 < r < rmax and 0 <  < 2 dependent power injections at buses i and j shown in
fig. .5.
14
Power Flow Analysis Using A Unified Power Flow Controller
ICAST-2017

Pi ,upfc  Pis  Pshunt  0.02rbseVi 2 sin 


(10)
1.02rvi v j bse sin i   j   

Qi ,upfc  Qis  rbseVi 2 cos  (11)

Pj ,upfc  Pjs  rvi v j bse sin i   j    (12)


Fig. 5. Equivalent power injections of series voltage
source Q j ,upfc  Q js  rvi v j bse cos i   j    (13)
B. Shunt Connected Voltage Source Converter
In UPFC, shunt connected voltage source is used III. IMPLEMENTATION OF UPFC MODEL IN
mainly to provide both real power, Pseries, which is POWER FLOW STUDIES
injected to the system through the series connected
voltage source, and the total losses within the Two imaginary PQ buses (bus i and bus j) are
UPFC. The total switching losses of the two created on line where UPFC is considered to be
converters is estimated to be about 2 % of the power located. In order to represent the model correctly,
transferred for thyristor based PWM converters series reactance Xse, is positioned between these
(Mohan, 1995). If the losses are to be included in two buses. When the position of UPFC on the
the real power injection of the shunt connected transmission line is changed, the line data, Z, should
voltage source at bus i, Pshunt is equal to 1.02 times be modified, depending on the location of the
the injected series real power Pseries through the UPFC. Here the UPFC is considered to position in
series connected voltage source to the system. the middle of line as illustrated in Fig. 7.

Pshunt  1.02 Pseries (8)

The complex power supplied by the series voltage


source converter is given as
*
S series  Vse I L  Pseries  jQseries
 vi v j bse r sin i   j     rbseVi 2 sin 
(9)
 jrvi v j bse cos i   j    
jrVi 2bse cos   jr 2bseVi 2

Fig. 6. Complete power injection model


The reactive power delivered or absorbed by shunt
converter is not considered in this model, but its
effect can be calculated and modeled as a separate
controllable shunt reactive source, The main
function of this reactive power is to maintain the
voltage level at bus i within acceptable limits, in this
case shunt converter functions as a static var
compensator. In view of the above explanations, we
assume that Qshunt=0. Consequently, the UPFC
power injection model is constructed from the series
connected voltage source model with the addition of
a power injection equivalent to PShunt + j0 to bus i as
shown in Fig. 6.

Where Pi ,upfc  jQi ,upfc and Pj ,upfc  jQ j ,upfc are


Fig.7. Modification of line data due to UPFC position
formulated as follows

15
Power Flow Analysis Using A Unified Power Flow Controller
ICAST-2017

IV. SIMULATION
1.11

1.1

VOLTAGE (P.U)
1.09
The performance of the UPFC injection model is 1.08

tested by carrying load flow studies on IEEE 14-bus 1.07

system (IEEE) embedded with UPFC by using 1.06

1.05

UPFC by using Newton Raphson method in 1.04

Matlab. The injection model is placed near bus 2 on 1.03

1.02

line 5, near to power generation stations as shown in 1.01


0 2 4 6 8 10 12 14
BUS NO
Fig. 8.
UPFC, power flows can be flexibly controlled by
UPFC. The simulations show that UPFC provides
independent control of active and reactive power
and improvement of voltage profile at the busses.

without UPFC with UPFC


Fig. 10. Voltage profile of 14 bus system

V. CONCLUSIONS

An injection modeling approach for power flow analysis


of power system with UPFC is studied. A steady state
mathematical model for the UPFC was proposed. The
Fig. 8. IEEE 14-Bus system embedded with UPFC proposed model can easily be incorporated in existing
power flow programs. In this paper the performance of
UPFC was investigated in controlling the flow of power
The UPFC has two control parameters r and  , the over the transmission line. Numerical results verify the
magnitude and phase of the injected voltage effectiveness of the model in terms of computational
respectively. Allowed iteration tolerance is taken as speed, accuracy and computing resources requirement. It
IE-5. First of all without any compensation, the was found that the UPFC regulates the voltage of the
electrical system is studied in order to determine the buses as well as regulates the active and reactive power
load flow in each of the transmission line, then after of the buses.
introducing the UPFC in the system and with
different UPFC parameters the voltage profile of all REFERENCES
buses, the transmitted active and reactive power of
1. L. Gyugyi, “Unified power flow control concept for
all lines are studied. All the results indicate good flexible AC transmission systems”, in IEEE Proc.-C,
convergence and high accuracy achieved by the Vol. 139, No. 4, pp. 323-331, July 1992.
proposed method. TABLE 1 shows the selected 2. M. Naroozian, and G. Anderson, “Power flow
results of the load flow analysis. Fig.8 and Fig. 9 control by use of controlled series
shows the graphical results related with simulation. components”,IEEE Trans. Power Delivery, Vol. 8,
Fig.9. Active power flows pp. 1420-1429, July 1993.
3. IEEE 14-bus test system data (Online). Available:
http://www.ee.washington.edu/
Comparing load flow solutions without and with research/pstca/pf14/pg tca14 bus.htm.
200 4. A. I. Nabavi-Niaki, M.R., “Steady-state and dynamic
models of unified power flow controller (UPFC) for
180 power system studies”, IEEE Transmissions on
Power Systems, Vol. 11, No. 4, pp. 1937-1943,
160
1996.
Power flow from bus 1-5

5. C.R.A. Fuerte-Esquivel E.; Ambriz-Perez H., “A


140
comprehensive Newton Raphson UPFC model for
the quadratic power flow solution of practical power
networks”, IEEE transactions on Power Systems,
120
Vol. 15, No.1, pp. 102-109, 2000.
100

16
80

60
1<0 0.91<10 0.96<15 0.94<20 0.95<25 0.95<30 0.92<35 0.91<40 0.89<45
UPFC variables
Power Flow Analysis Using A Unified Power Flow Controller
ICAST-2017

6. C.R.A. Fuerte-Esquivel E.; “Unified power flow 7. Ned Mohan, Power Electronics: Converters,
controller: a critical comparison of Newton-Raphson Applications, and design, New York: Wiley, 1995
UPFC algorithms in power flow studies”, IEEE
proceedings on Generation, Transmission and
Distribution, Vol. 144, No. 5, pp. 437-444, 1997.

Line power flow in p.u


Bus no without UPFC With UPFC
r=0, θ=0° r=0.95, θ =25° r=0.91, θ =40° r=0.89, θ =45°
1-2 1.568-i0.2335 0.84439-i0.0482 0.3860-i0.6252 0.3679-i0.6198
1-5 0.7556-i0.2806 1.348-i0.0271 1.740+i0.0048 0.1893+i 0.0621
2-3 0.7280+i0.0121 1.071-i0.0104 1.315+i0.1920 1.396+i10.9231
2-4 0.5586-i0.6517 1.273-i0.1497 1.814+i0.0301 1.972 +i0.0680
4-5 -0.61914-i0.1629 0.2879-i0.2688 0.9279-i0.4219 1.111-i0.4530
Total active
0.1334 0.0187 0.0839 0.2221
power loss
Total
reactive 0.0534 0.3265 0.1560 0.2118
power loss

Table I. Load Flow Results Of Ieee 14-Bus System Without And With Upfc

17
Power Flow Analysis Using A Unified Power Flow Controller
ICAST-2017

18
On Board Charging Interface for Vehicle to Grid (V2G) in Smart Grid Environment
ICAST-2017

On Board Charging Interface for Vehicle to Grid (V2G) in Smart


Grid Environment
Dr.Santoshkumar1, Kena Likassa Nefabas 2, Hinsermu Alhemayu3
1 ,2,3
School of Electrical Engineering and Computing,
Adama Science & Technology University, Adama, Ethiopia
E-mail: 1antoshkumar777@yahoo.com,2 kena0912@gmail.com,3 hialex98@gmail.com

Abstract — The developments in the power grid and advancements in vehicle technology made
Vehicle to Grid (V2G) a reality. The concept of connecting a group of electric vehicles (EV) to the grid
for power transaction is called V2G. Vehicle to Grid is used for peak shaving, valley filling, meet the
time varying load demand and helps in maintaining grid stability for a shorter duration of time.
electric vehicles need a bidirectional charger to either draw or inject power to the grid. In this abstract
On board charging for V2G in Smart Grid environment is proposed. In this abstract an architectural
framework of On-board V2G Integrator is proposed. A single phase on board charger with low
complexity control scheme is presented for EV power transaction. The system designed for voltage
rating of 230V and the simulations is carried out in MATLAB/SIMULINK and obtained results
comply with IEEE 1547 standard.
Key words- On board charger, EV, V2G, Smart Grid
state error and also the limited bandwidth hinders
I. INTRODUCTION the response time. The DC-DC converter employed
has both the capability of stepping up (boost) the
The concept of connecting a group of electric DC link voltage during V2G operation and stepping
vehicles (EV) to the grid for power transaction is down (buck) the DC link voltage during G2V
called Vehicle to Grid (V2G). V2G plays a operation in comparison to battery voltage.
prominent role in fulfilling the grid requirements
and meet the load demand. EVs also support the
ancillary services like load leveling, voltage
regulation, frequency regulation and balancing. EVs
need a bidirectional charger to sell or buy power
from the grid. Farther the bidirectional charger has
the direct current (DC) link capacitor which is
inherently able to provide the reactive power to
support the power grid. The state of charge (SOC)
of the EV battery plays a key role in V2G operation
and promotes the concept of Vehicle-to-Home
(V2H), Vehicle-to-Vehicle (V2V) and V2G (Liu et
al., 2013; Pinto et al, 2014). The architectural and
conceptual framework of V2G is shown in Fig. 1.
The EV can be connected to the home grid or other
interface using the On-Board or Off-Board
bidirectional charger or V2G integrators. In this Fig.1. Conceptual framework of V2G
abstract an architectural framework of On-board
V2G integrator is proposed. A single phase on II. PROPOSED ON BOARD V2G INTEGRATOR
board charger with low complexity control scheme
is presented for EV power transaction. The system Fig.2 shows the topology diagram of the proposed
designed is also suitable for V2H as the charger system consisting of cascaded power converters. It
considered is designed for voltage rating of 230 V. consists of two power stages. The first stage is
The DC-AC bidirectional converter is made to bidirectional DC-DC converter and the second stage
operate at unity power factor injecting or drawing is a bidirectional DC-AC converter. The direction of
the grid current with low harmonic distortion. The power flow is dictated by the operating mode. The
control scheme employed consists of a PR active power flows from the batteries to the dc link
controller for grid current control as an alternative and then to the grid during V2G mode of operation.
to the widely used PI controller for both V2G and The power flow from and to the grid is processed
G2V operation (Teodorescuet al, 2006). The low gain using two stage cascaded converters consisting of a
of PI controller at grid frequency leads to steady bidirectional DC-DC and DC-AC converters. The

19
On Board Charging Interface for Vehicle to Grid (V2G) in Smart Grid Environment
ICAST-2017
LCL filter is used as an interface between DC-AC
converter and the grid to attenuate the grid current Table I. Specification of the Proposed On Board Charger
harmonics. A proportional-resonant (PR) controller
Parameter Value Unit
is employed for the control of grid current and to
enable the unity power factor operation of the DC- Grid voltage
230 V
AC converter. (RMS)
Grid frequency 50 Hz
Maximum input
35 A
current (RMS)
Maximum input
Fig.2.Topology of the proposed on board charger. 8 KVA
power
The setbacks associated with conventional
proportional integral (PI) controller for single phase Output voltage 190-
system is elevated by employing PR controller. The V
range (DC) 270
power electronic interface for the power transaction
from V2G and G2V may involve one or more Maximum output
stages. Most of the EV chargers are of conductive 30 A
type, where there is an electrical contact between current (DC)
the vehicle and the utility. Bidirectional converter
topologies are to be used chargers for V2G Inductor L1 0.3 mH
operation. In addition to traction batteries the
auxiliary battery used for lighting, wipers needs to Inductor L 2 0.5 mH
be charged. A reconfigurable battery charger in
which the auxiliary battery is charged by the main Capacitor C f 2 F
traction battery is proposed in (Yilmaz & Krein, 2013
;Teodorescuet al, 2006). Capacitor C b 800 F

III. VEHICLE TO GRID MODE OF OPERATION Capacitor C dc 1000 F


In this mode of operation the DC-AC bidirectional
converter injects the sinusoidal current into the grid Switching
in phase with the grid voltage. The reference current
generated by the PI controller 1 is compared with frequency of DC- 10 kHz
the actual grid current and the error is processed AC converter
through the PR controller. The DC-DC converter
acts as a boost converter by stepping up the battery Switching
voltage and draws a constant current from the
battery. The specification of the various parameters frequency of DC- 25 kHz
of the system and the controller parameters are
DC converter
shown in Table I and II respectively.

A. Three level bidirectional DC-AC converter Table II. Parameters of various controllers
control
KP KI
In grid connected operation, the converter operates
in a current controlled mode as the voltage across it PI controller 1 0.3 2
is maintained by the infinite grid. Thus, for proper
operation, synchronization with the grid voltage is PI controller 2 0.75 5
mandatory. This is achieved using a phase locked
loop (PLL) which generates an in phase component PR controller 3 10000
of the grid voltage and eliminates any harmonics
present.

20
On Board Charging Interface for Vehicle to Grid (V2G) in Smart Grid Environment
ICAST-2017
An LCL filter acting as an interface between the gain at fundamental frequency, H 1  j   would have
DC-AC converter and the grid reduces the harmonic
a unity gain. The control structure of a PR controller
distortion in the grid current being injected. As the
is given in equation (4).
filter is of order three, the amount of attenuation
K i  2    o  s
provided over high frequency range is more even Gc  s   K p 
with a smaller value of passive components. s  2    o  s  02
2
(4)
However the filter poses a significant problem due
to its low or zero impedance at the resonant where o  2 fo , K i is the fundamental
frequencies makes the design of current controller harmonic gain, and  is the damping factor.
complicated. A general approach to damp the
resonance oscillations is to add a damping resistor The control block diagram of the bidirectional DC-
in series with the filter capacitor. Even though the AC & DC-DC converter is shown in Fig.3.The gain
method seems to be effective in stabilizing the of around 82 dB at the grid frequency (50 Hz) leads
overall filter characteristics, it suffers from the to zero steady state error in tracking of the reference
increased power loss. current generated. The converter DC link voltage
should be greater than the peak value of the grid
As an alternative to passive damping, approach to voltage, so that the power can be transferred from
introduce the same effect using feedback of the EVs to the grid. Assuming that the sinusoid current
parameters which can serve as a damping term is being pumped into the grid, the DC link voltage
known as active damping. For an application with a will have a second harmonic component. This leads
stiff grid, a passive damping method is often to control system instability if it is not filtered out
preferred for its simplicity and low cost. The control before using for control purpose. Thus, band stop
of DC-AC converter is achieved through two loops
filter (BSF) is used to remove 100 Hz component in
with outer loop as voltage control loop and with
Vdc. The unipolar pulse width modulation scheme
inner loop as a current control loop. Various
controllers like stationary frame control, dq frame is employed for gating the power switches.
control and abc frame controllers are employed to
maintain the grid current to be sinusoidal with lower
harmonic distortion. In this abstract a proportional
resonant (PR) compensator is used to track a
sinusoidal current reference signal with zero steady
state error as the controller introduces an infinite
gain theoretically at the grid frequency.
 s 2 L1Cd  sCd Rd  1 
I g  s   Gp  s   Vg  Vinv 
 sCd Rd  1 
(1)
AS, the magnitude and phase response of
Fig.3. Control block diagram of the bidirectional DC-AC
s L1Cd  sCd Rd  1
2
& DC-DC converter.
are 0 dB and 0o at the
sCd Rd  1
B. Bidirectional DC-DC Converter Control
fundamental frequency of the grid. Therefore,
equation (1) can be simplified to the equation The DC link voltage is maintained constant by
the DC-AC converter and the DC-DC converter
I g  s   Gc  s  Vinv  Vg  acts as a buck converter during this mode
(2)
(G2V). As per the recommendation by the EV
The relationship between the input and the output of manufacturer generally a constant current
the current loop can be derived as:
charging is done first till the battery voltage
I g  s   H1  s   I g*  s   H 2  s  Vg  s  reaches the recommended maximum value and
(3) then followed by a constant voltage charging. A
To successfully track the ig* t  signal without proportional integral controller (PI) is used for
regulating both the DC link voltage and the
steady state errors, the magnitude of H 1  j   has current (charging/discharging) current from the
to be equal to 1 at the fundamental frequency of EV battery.
ig* t  . Thus, it is clear that if Gc  j   has infinite
21
On Board Charging Interface for Vehicle to Grid (V2G) in Smart Grid Environment
ICAST-2017
IV. RESULTS AND DISCUSSION
To demonstrate and validate the feasibility of on
board Integrator, simulation was carried out using
MATLAB/SIMULINK. Fig. 4, Fig.5 and Fig.6
show the regulated DC link voltage, battery voltage
and the battery current during V2G operation. The
bidirectional DC-DC converter acts as a boost
converter during this mode. Fig.4 shows the grid
voltage and the sinusoidal current being injected to
the grid. It can be seen that the current injected is in
phase with the grid voltage ensuring the unity power
factor operation. The battery is supplied with a
constant active power (P = 6 kW) and zero reactive Fig.5. Active and reactive power.
power (Q = 0) as shown in Fig.5. The harmonic
level in the grid current being drawn is found to be
2%. The harmonic spectrum of the grid current is
shown in Fig.6.

VI. CONCLUSION
A low complexity on-board charger for EVs has
been presented. The proposed system allows the
integration of EV battery to the grid through power
processing stages. The power flows from vehicle
battery to grid during V2G mode of operation. The
grid current being injected/absorbed is maintained
to have a power factor of unity and low harmonic
distortion. Design details of the current controller
are discussed and simulation results are presented.
Simulation is carried out using the MATLAB /
SIMULINK platform and the result showing the Fig.6. Harmonic spectrum of grid current (2% THD).
harmonic analysis of the grid current is found to be
within the prescribed power quality limits of IEEE. REFERENCES
1. Liu, K., Chau, D., Wu., S. Gao.(2013). Opportunities
and challenges of vehicle-to-home, vehicle-to-
vehicle, and vehicle-to-grid technologies.
Proceedings of the IEEE, 101(2013), 2409– 2427.
2. J. Pinto., V. Monteiro., H. Gon¸calves., J. L.
Afonso.(2014). Onboard reconfigurable battery
charger for electric vehicles with traction-to-
auxiliary mode. IEEE Transactions on Vehicular
Technology, 63(2014), 1104–1116.
3. M. Yilmaz., P. T. Krein.(2013). Review of the
impact of vehicle-to-grid technologies on
distribution systems and utility interfaces. IEEE
Transactions on Power Electronics, 28(2013),
5673–5689.
Fig.4. Grid voltage and current.
4. R. Teodorescu., F. Blaabjerg., M. Liserre., P. C.
Loh.(2006). Proportional-resonant controllers and
filters for grid-connected voltage-source converters.
IEE Proceedings on Electric Power Applications,
153(2006), 750–762.
5. K. Thirugnanam., T. E. R. Joy., M. Singh., P.
Kumar.(2014). Modeling and control of contactless
based smart charging station in V2G scenario. IEEE
Transactions on Smart Grid 5(2014), 337–348.
6. M. A. Khan., I. Husain., Y. Sozer.(2012). A bi-
directional dc-dc converter with overlapping input
22
On Board Charging Interface for Vehicle to Grid (V2G) in Smart Grid Environment
ICAST-2017
and output voltage ranges and vehicle to grid energy
transfer capability. Proceedings of IEEE
International Electric Vehicle Conference (IEVC),
2012, 1–7.
7. S. Kumar., U. R. Yaragatti., S.
Manasani.(2014).Modeling and architectural frame
work of off-board V2G integrator for smart grid.
International Journal of Renewable Energy
Research (IJRER) 4(2014), 831–836.

23
On Board Charging Interface for Vehicle to Grid (V2G) in Smart Grid Environment
ICAST-2017

24
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

Phase shifted carrier PWM based multilevel converters for


1.5 MW wind turbines
P.Palanivel1, Tafese Asrat2, Kena Likassa3, Hinsermu Alemayehu4, Satish sukhavasi5
Electrical Power and Control Engineering Program Adama Science and Technology University
Adama, Ethiopia
drpalanivelres@gmail.com

Abstract — The development of wind energy technologies had several significant trends in the past
fifteen years, like the size and capacity of the wind turbines grow rapidly, and making the price also
continuously reduced. With stricter power grid integration requirements, the power electronics
converters are changing from partial rated power to full rated power, and playing a more and more
important role in the whole generation system. As it is troublesome to connect only one power
semiconductor switch directly, a multilevel power converter structure has been introduced. To obtain a
quality output voltage with minimum amount of ripple content, they require high switching frequency
along with advance pulse width modulation strategies. As a result, a multilevel power inverter
structure has been introduced as an alternative in high power applications. In this paper, phase shifted
carrier pulse width modulation based multilevel converter configuration for 1.5 MW wind turbines
are proposed, designed and compared. Here performance analyses the phase shifted carrier pulse
width modulation techniques applied to multilevel converters. Concepts of pulse width modulation
techniques are presented and their performances are analyzed. Due to harmonics present in the
inverter, the total harmonic distortion is very high and output voltage is very low. To reduce total
harmonic distortion value and to maximize output voltage, phase shifted carrier pulse width
modulation techniques are proposed. Finally, to achieve 3.84% THD and enhance the 5% output
voltage using phase shifted carrier based multilevel inverters for 1.5MW wind turbine.
Key words- Wind turbines, phase shifted carrier pulse width modulation, multilevel inverter
In this paper, phase shifted carrier pulse width
I. INTRODUCTION modulation based cascaded multilevel converter
Wind energy is the fastest growing energy configuration for 1.5 MW wind turbines are
technology in terms of percentage of yearly growth proposed, designed and compared. Here
of installed capacity per technology source. As wind performance analyses the phase shifted carrier pulse
energy is increasingly integrated into power width modulation techniques applied to multilevel
systems, the stability of already existing power converters. To reduce total harmonic distortion
systems is becoming a concern of utmost value and to maximize output voltage, phase shifted
importance. Also, network operators have to ensure carrier pulse width modulation techniques are
that consumer power quality is not deteriorated proposed. The evaluation criteria will mainly aim
(Blaabjerg et al., 2004). Hence, the Total Harmonic at the utilization and thermal performances of power
Distortion (THD) should be kept as low as possible, switching devices.
improving the quality of the energy injected into the
electric grid. The new technical challenges II. MULTILEVEL CONVERTERS FOR WIND
emerging due to increased wind power penetration, TURBINE
dynamic stability and power quality imply research
As the interface between the wind turbine generator
of more realistic and accurate models for wind
and the power grid, the wind power converter has to
energy systems (Chen et al., 2009). satisfy the requirements on the both sides, as
Power-electronic converters have been developed summarized in Fig. 1.
for integrating wind power with the electric grid
(Blaabjerg et al., 2004). The use of power-electronic
converters allows not only for variable-speed
operation of a wind turbine, but also for
enhancement on power extraction. In a recent
overview of different wind generator systems, it is
shown that variable-speed conceptions equipped
with power-electronic converts will continue to
dominate and be very promising technologies for
large wind farms (Polinder et al., 2006). Fig.1. Wind power converter system

25
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
For the generator side: The current flowing in the III. DESIGN CRITERIA
generator’s stator should be controlled to adjust the
In order to conduct evaluation of each converter
rotating speed. This will contribute to the active
candidate, the parameters of the generator and basic
power balance during grid faults and help to extract
design for converters are needed.
the maximum power from the wind turbines.
Moreover, the converter should have the ability to
handle the variable fundamental frequency and
Table.I. Design Parameter For 1.5 Mw Wind Turbine
voltage amplitude of the generator’s output.
WT parameter Value
From the grid side: The converter must comply with
Wind power capacity 1.5MW
the grid codes regardless of the wind speed. That
means it should have the ability to control the Generator voltage 2200V
injected/absorbed reactive power Q, and perform a Generator current 852A
fast active power P response. The fundamental Multilevel inverter rating 2MVA
frequency as well as voltage amplitude on the grid Input voltage 3110V (DC)
side should be almost fixed under normal operation, Input current 1204A (DC)
and the total harmonic distortion of the current must Fundamental frequency 50Hz
be maintained at a low level. Switching frequency 5000Hz
IGBT voltage rating 3.3KV
Inherently, the converter needs to satisfy both the
generator side and grid side requirements with a IGBT current rating 1.5KA
cost effective and easy maintenance solution. That Regarding the parameters of the converter, all of the
requires a high power density, reliability, power switching devices have the commutated
modularity of the whole converter system. voltage at 3.3 kV in order to utilize the available and
Moreover, the wind power converter may need the dominant 4.5 kV high-power IGCT/IGBT on the
ability to store the energy, and boost up the voltage market (Blaabjerg et al., 1995). For simplicity of
from generator side to the grid side. According to analysis, the power grid is considered as three 20
the overall demands, multilevel converter kV/50 Hz, ideal AC voltage sources, the resistance
configurations for 1.5 MW wind turbines are in the generator and the cables is not taken into
proposed. account, DC bus capacitance is assumed high, and
Generally, multilevel converters can be classified in the transformers are assumed ideal.
to three categories: Neutral point, diode clamped A. Design of grid side converter
and cascaded multilevel structures (Rodriguez et al.,
According to the commutated voltage of power
2010). In order to get a cost effective design and
devices (3.3 kV), the DC bus and maximum output
less complexity cascaded multilevel converters are voltage of each configuration can be determined.
better performances (Bruckner et al., 2005). 5L-HB The equivalent switching frequency fs of the grid
BTB: This configuration is composed of two back- side converter is designed at 5000 Hz in order to get
to-back five level H-bridge converters (Holtz & an acceptable switching loss of the power devices.
Oikonomou , 2010). In this wind power system The output filter inductance is designed to limit the
generator side five level converter and grid side also maximum current ripple to 25% of the rated
five level converters can be used (Kouro et al., maximum current amplitude, and the filter
2010). 5L-HB BTB can achieve five-level output capacitance is not taken into account.
voltage and doubled voltage amplitude.
Table.II. Parameters Of Grid Side Converter At
Different Wind Speeds
Wind speedVW 3 6 9 11
(m/s)
Generator 0.4 0.8 1.2 1.5
powerPG (MW)
Primary side 2200V
voltage (Vrms)
Fundamental 50
frequency (Fg)
Switching 5000
frequency (Fs)

B. Design of generator side converter


Fig.2. Five level HB back to back converter The equivalent switching frequency of the generator
side converter is designed as 10 times of the
26
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
generator’s maximum fundamental frequency fe. For The operating rules for PSC PWM when the number
simplicity, filters on the generator side are not of level n = 5 is given below:
considered. The constant stator voltage control is
• The n – 1 = 4 carrier waveforms are arranged. The
used when the wind speed is above 11 m/s, and the
carriers between the full bridge inverters are phase
maximum torque control is used when the wind
shifted 90º.
speed is below 11 m/s.
• The converter switches to + Vdc when the
Table.III. Parameters Of Generator Side Converter At
reference is greater than all the carrier waveforms.
Different Wind Speeds
• The converter switches to Vdc / 2 when the
Wind speedVW 3 6 9 11
reference is less than the uppermost carrier
(m/s)
waveform and greater than all other carriers.
Generator 0.2 0.6 0.9 1.2
powerPG (MW) • The converter switches to 0 when the reference is
Induced voltage 1800
(Vrms)
Fundamental 50
frequency (Fg)
Switching 800
frequency (Fs)

IV. PHASE SHIFTED CARRIER PULSE WIDTH


MODULATION (PSC PWM)

less than the two uppermost carrier waveform and


greater than two lowermost carriers.
• The converter switches to - Vdc / 2 when the
reference is greater than the lowermost carrier
waveform and lesser than all other carriers.
Fig.4. PSC PWM signal generation
• The converter switches to -Vdc when the reference
Fig.3. Phase shifted carrier PWM is lesser than all the carrier waveforms.
Fig.2 shows the Phase shifted carrier pulse width Phase shifting for the carrier is given by,
modulation. Each cell is modulated independently
using sinusoidal unipolar pulse width modulation ( K  1) 
and bipolar pulse width modulation respectively, Pcr 
n (1)
providing an even power distribution among the
cells The reference voltage is continuously Where, K is the Kth bridge.
compared with each of the shifted carrier signals.
Each cell is modulated independently using the n is the number of series connected single phase
PWM, which provides an even power distribution inverter.
among the cells. A carrier phase shift of 180°/m for n = (L-1)/2 (2)
the cascaded inverter is introduced across the cells
to generate a stepped multilevel output waveform Where, L is the number of switched DC levels that
with lower distortion, where ‘m’ is the number of can be achieved in each phase Leg.
full bridge inverters in a multilevel phase leg. The average output voltage for a phase shifted pulse
Fig.4 shows the PSC PWM signal generation. For n- width modulation to a particular power cell ‘i’ is
level converter, (n-1) phase shifted carrier signals given by
are generated. The carriers between the full bridge Voi = 1/Tcr . ʃVoi(t) dt (3)
inverters are phase shift 180º/m. If the reference is
greater than carrier signal, then the active device Voi = Ton/Tcr . Vdc (4)
corresponding to that carrier is switched off. Voi = V (5)

27
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
Where, Voi is the output voltage of cell i, and Ton is same output level. The PSCPWM frequency spectrum is
the time interval, determined by the comparison
between the reference and the carrier signals.
The three phase sinusoidal modulating signals are
generated by using a phase shift oscillator. This
signal is compared with (n-1) phase shifted carrier
waves and PWM pulses are generated. These PWM
pulses are applied to three phase five level inverter.

shown in Fig.7. In frequency spectrum the switching


frequency is 5 KHz with fundamental frequency 50 Hz.
The output voltage obtained by
Fig.7. PSCPWM based five level converter frequency
spectrum
PSCPWM is about 3212V for input voltage of
1550V from each source. As switching frequency is
5 KHz and fundamental frequency is 50Hz so
harmonic order is about 100 which is shown in
Fig.8. The THD value is about 3.84%.

Fig. 5. Simulation output voltage for five level converter


wind turbine
Fig.5 shows the simulation output voltage for
multilevel inverter. The output voltage is 3212V.
But input voltage is only 3110V. Here output
voltage is enhanced for nearly 5%. So, phase shifted

Fig.8. PSCPWM based five level converter harmonic


spectrum
It is observed that the PSC PWM gives better result
compared to the other methods. Here, the PSCPWM
strategy reduces the THD and enhances the output
voltage. The output voltage Vac is maintained
3212V. The THD value is 3.84%.
VI. CONCLUSION
The 1.5 MW wind power converter, 5L-HB BTB
shows excellent performances on the grid side and
generator side. The phase shifted carrier PWM
technique gives superior performance in terms of
enhanced output voltage and reduced THD when
carrier PWM is used to enhance the output voltage. compared to other techniques. This results analyzed
This technique used to wind power converter by simulation using MATLAB simulink and their
systems. performance analyzed by implementing FPGA
SPARTAN-3 processor. The results are obtained
Fig. 6. Hardware output voltage for five level converter, from experimental work which is almost similar to
wind turbine the simulation work.
The hardware output voltage shown in Fig.6. The
hardware and simulation output voltage more or less

28
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
REFERENCES
1. F. Blaabjerg, Z. Chen, S.B. Kjaer. (2004) Power
Electronics as Efficient Interface in Dispersed Power
Generation Systems, IEEE Transactions on Power
Electronics, 19 (4), 1184-1194.

2. Z. Chen, J.M. Guerrero, F. Blaabjerg (2009) A


Review of the State of the Art of Power Electronics
for Wind Turbines, IEEE Transactions on Power
Electronics, 24(8), 1859-1875.

3. S. Kouro, M. Malinowski, K. Gopakumar, J. Pou, L.


G. Franquelo, B. Wu, J. Rodriguez, M. A. Perez, J. I.
Leon. (2010) Recent Advances and Industrial
Applications of Multilevel Converters, IEEE
Transactions on Power Electronics, 57(8), 2553 –
2580.

4. D. Krug, S. Bernet, S. S. Fazel, K. Jalili, M.


Malinowski. (2007) Comparison of 2.3-kV Medium-
Voltage Multilevel Converters for Industrial
Medium-Voltage Drives, IEEE Transactions on
Industrial Electronics, 54(6), 2979-2992.

5. J. Rodriguez, S. Bernet, P. K. Steimer, I. E. Lizama.


(2010) A Survey on Neutral-Point-Clamped
Inverters, IEEE Transactions on Industrial
Electronics, 57(7), 2219-2230.

6. T. Bruckner, S. Bernet, H. Guldner. (2005) The


active NPC converter and its loss-balancing control,
IEEE Transactions on Industrial Electronics,. 52(3),
855-868.

7. J. Holtz, N. Oikonomou. (2010) Optimal Control of


a Dual Three-Level Inverter System for Medium-
Voltage Drives, IEEE Transactions on Industrial
Applications, 46(3) 1034-1041.

8. H.Polinder, F.F.A.van der Pijl, G.-J.de Vilder,


P.J.Tavne.r (2006) Comparison of direct-drive and
geared generator concepts for wind turbines, IEEE
Transactions on Energy Conversion, 21(3) 725-733.

9. F. Blaabjerg, U. Jaeger, S. Munk-Nielsen and J.


Pedersen. (1995) Power Losses in PWM-VSI Inverter
Using NPT or PT IGBT Devices, IEEE Transactions
on Power Electronics, 10(3), 358–367

29
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

30
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

Performance Analysis of Multiple Antenna Based Blind


Spectrum Sensing Techniques for Cognitive Radio Networks
Amare Kassaw1, Dr.Ing.Dereje Hailemariam2, Prof. Dr.-Ing. Abdelhak M. Zoubir3
1
Bahir Dar Institute of Technology, Bahir Dar University, 2Addis Ababa Institute of Technology, Addis Ababa
University,
3
TechnicalUniversity of Darmstadt
Email: amex2121@gmail.com, derejehmr@gmail.com,zoubir@spg.tu-darmstadt.de

Abstract-- This paper presents the state-of the-art on blind signal detection and spectrum sensing
techniques that are used to solve some of the major practical issues and design challenges in signal
detection and spectrum sensing. Specifically, the paper reviews multiple-antenna based blind
spectrum sensing algorithms that work without a-priori knowledge of the primary user's signaling
scheme, the propagation channels, and the variance of the noise seen at the secondary users. The
performance of these blind spectrum sensing algorithms are analysed through simulation for spatio-
temporal correlation condition. The results show that the performance of these detectors are very
good and they are blind(robust) that work well without a priori knowledge of the PU signal, noise
variance and channel gains.
Keywords:Dynamic Spectrum Access, Cognitive Radio, Spectrum Sensing
and CR users, it is often compromised by factors
such as hidden terminal problem,multipath fading,
I. INTRODUCTION
shadowing and receiver uncertainty, and other
Nowadays, the need for higher data rates is issues (Yucek and Arslan, 2009, Akylidiz et.al,
increasing as a result of the rapid evolutions of new 2011; Zeng et al, 2010) as shown in Figure 1.To
wireless services and applications. However, the mitigate the impact of these issues, cooperative
radio frequency spectrum allocation and spectrum sensinghas been seen as an effective
management is still based on the traditional method toimprove the detection performance by
techniques from the early day of wireless exploiting spatialdiversity. The main idea of
communication (Mitola and Maguire,1999). cooperative spectrum sensing is to enhance the
sensing performance by exploiting the spatial
This static frequency allocation schemes cannot
diversity in the observations of spatially distributed
meet the requirements of these increasing demands
CR users. By cooperation, CR users can share their
(Mitola and Maguire,1999).This inefficiency in the
sensing information to make a combined decision
spectrum usage and allocation necessitates a new
more accurate than the individual decisions
communication paradigm to effectively explore and
(Hykin,2005). Cooperative spectrum sensing is used
exploit the radio frequency spectrum opportunities.
to avoid hidden terminal problems, improve
Hence, Dynamic Spectrum Access (DSA) is
detection performance and relaxed sensitivity
proposed in(Mitola and Maguire, 1999; Hykin,
requirements, and reduce the sensing time which
2005) which allows wireless devices to operate
intern used to improve the throughput performance
opportunistically in spectrum opportunities by
of the network (Akylidiz,et.al,2011,Zeng,et al,2010).
controlling the harmful interference and overheads
incurred on the licensed users.
The key enabling technology for DSA is the
cognitive radio (CR) that is built on a software
defined radio which involves dynamic spectrum
access in which a set of unlicensed users occupy
spectrum holes without causing significant
degradation of performance to the incumbent
users(Mitola and Maguire,1999,Hykin,2005). The
first fundamental step in CR networks is spectrum
sensing which aims to learn the radio environment
in order to adapt the CR transmission. Thus, reliable
and efficient signal detection and spectrum sensing
techniques are needed in order to detect and explore
the presence of radio frequency spectrum holes and Fig. 1. Multipath fading, shadowing and receiver
vacant spaces. Even if the detection performance in uncertainty in spectrum sensing (Akylidiz, et.al,
spectrum sensing is very crucial for both licensed 2014).

31
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
Thus, facing all these design challenges, design and the transmitted signal (Zeng and Liang,2009, Lim,
development of robust signal detection and et.al,2008). For this, we consider the system model
spectrum sensing techniques that assist to mitigate (shown in Figure 2) of a cognitive radio receiver or
some or all of the above problems has been of great detector with M≥1 antennas that receive PU signal
interest in the study of cognitive radio via multipath fading channels. It is assumed that a
networks. This paper specifically reviewed and central unit is available for combining and
analyzed blind(robust) multiple antenna based processing the signals from all the antennas.
cooperative spectrum sensing algorithms that are
Spectrum sensing is a hypothesis testing problem
capable to mitigate the noise uncertainty problem
based on binary hypotheses (H0; H1), where H0
and work at hostile wireless channel
indicates the primary signal is absent; and H1
conditions(Zeng and Liang,2008,Lim, et.al,2008).
indicates the primary signal is present. Based on the
These spectrum sensing algorithms work well
above system model, the received signal at
without a-priori knowledge of the PU signaling
antenna/receiver i can be formulated as;
scheme, the wireless channel, and the noise
variances.
The rest of the paper is organized as follows: In
section II, the system model for multiple antenna
where i=1,2,…, M, wi(n) is noise at the receive
based cooperative spectrum sensing is shown.
antennas, si(n) is the received source signal at
Section III formulates the test statistics for
receiver i , assuming that it is the received signal
Eigenvalue based detection algorithms for multiple
after passing through the fading and multipath
antenna based cooperative spectrum sensing.
propagation channels, it can be given by;
Section IV shows simulation results and analysis.
Finally conclusion is given in Section V.

II. SYSTEM MODEL FOR COOPERATIVE


SPECTRUM SENSING Where sp(n) shows the transmitted signal from
Cooperative wireless communication is a new primary user, hi(l) denotes the propagation channel
communication paradigm that brings significant coefficients from the primary user to the ith receiver
capacity and multiplexing gain improvement in antenna and Ki is the channel order for ith receiver
wireless networks (Yucek and Arslan,2009, antenna.
Akylidiz,et.al,2011,Zeng,et al,2010). It also realizes In this work, we assume that the temporal
a new form of space diversity to combat the critical correlation is WSS. Hence, consecutive received
problems of severe wireless channel fading and samples of the (M ×1) vector X(n) at M antennas of
shadowing effects (Akylidiz,et.al, 2011). In this a user are temporally correlated due to
section, the model for practical scenario in which oversampling at the receiver, where X(n)=
the detection of the presence of a primary user (PU) [X1(n);X2(n); … ; XM(N)]. To process the problem,
by using multiple antenna assisted cognitive radio we need the distribution of the vector sample X(n),
under spatio-temporal correlated conditions is and it is assumed to bezero-mean complex
shown. Gaussian. This assumption is particularly
reasonable for primary user network that use
orthogonal frequency division multiplexing
(OFDM) as modulation format (Huang, 2013,
Ramirez, et.al,2008). It is also assumed that the SU
collects consecutive samples of vector X(n). In
order to apply this space-time processing, the
received vector samples are stacked in to L- blocks,
and each block shall consist of N- vector samples to
yield the following vector blocks;

Fig. 2. System model for multiple antenna cooperative


spectrum sensing Based on this stacked matrix, let us define
containing
The spatial correlation may be due to the proximity N-time samples of the ith time
or improper separation between the receiving block, ,
antennas; and the temporal correlation may be due and . Now, by
to the presence of temporal dispersive channel, formulating the covariance matrix as
oversampling at the receiver, or time correlation of v=0,1, the distributions of
32
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
observations under each hypothesis can be
formulated as ;
Where determinant and inverse of the covariance
matrix is given by (Kay,1998)
By using these distributions of the received signal
and noise, different approaches are formulated to
determine the decision test statistics, and the
detection performance metrics of spectrum sensing
in cognitive radio networks. The following section
formulates the test statistics for blind signal
detection algorithms based on Covariance and The GLRT requires the maximum likelihood
Eigenvalue approaches such as:Blindly Combined estimates of unknown parameters under both
Energy Detection (BCED), Maximum-Minimum hypotheses. Thus, we find the maximum likelihood
Eigenvalue Detection (MMED), the Spherical Test estimations of equation (7) with respect to the
Detector (STD), and the John's Detector (JD). parameters as;

III. TEST STATISTICS FOR BLIND SPECTRUM


SENSING TECHNIQUES
1. Blindly Combined Energy Detection(Bced)
The test statistics for BCED does not need any
information of the source signal and the channel a
priori, and it works well for correlated signals. Most
importantly, it does not need noise power estimation Then, by applying Eigenvalue decomposition
and overcomes susceptibility to noise uncertainty in principle (Kay,1998) on the covariance matrix R
energy detection (Zeng and Liang,2008, =UVUT, where U is the orthogonalization matrix
Lim,et.al,2008). To formulate the test statistics, this (Eigenvector), and V = diag(λ1; λ2; …; λML), we get
approach uses the Generalized Likelihood Ratio the maximum likelihood estimation of the three
Test(GLRT) (Lim,et.al 2008). For cognitive radio unknown parameters as follows
detector with M≥1 antennas, N-time samples, with (Lim,et.al,2008,Zeng and Liang,2008);
over sampling rate/smoothing factor (L), the general
form of the GLRT to the received vector X(n) can
be written as;

Where the PDF's of the observation vector X(n)


under each hypothesis is given by;

Where Um denotes the eigenvector corresponding


to the largest eigenvalue. From analysis, it can be
shown that (Lim, et.al,2008, Kay,1998) the
maximum likelihood estimator of the channel vector
h will be the eigenvectors of the sample covariance
matrix that is where λ is the
Where tr(.) denotes trace of the matrix,
and is the corresponding eigenvalue of the sample covariance
covariance matrix under the hypothesis . Taking matrix. For normalized channel vector , it
the logarithms of the PDF’s , we get ; can be also shown that (Lim, et.al,2008, Kay,1998)
the maximum likelihood estimation h is the
eigenvector corresponding to the maximum
eigenvalue of the sample covariance matrix that is
denoted by . Now, substituting these
maximum likelihood estimations to equation (7), we
get the log-likelihood ratio equation as;

33
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

4. Set the eigenvalue of the sample covariance


matrix in descending order as:

5. Calculate the detection threshold from equation


(28)
6. Calculate the test statistics as :
Where 7. Decision: If TBCED>τ, the primary user signal is
existed (decide in favour of H1), otherwise the
If the eigenvalue is sorted primary user is off.
as , it can be
seen that the supremum and infinimum of the
2. The Maximum Minimum Eigenvalue Detection
expression is given by . Now (MMED)
as the above LLR is an increasing function of in
the interval the GLR test in equation (14) To detect the presence of the primary user, the
can be simplified as;
test statistics of this method is developed from
the eigenvalue of the covariance matrices using
the ratio from the maximum Eigenvalue to the
minimum Eigenvalue (Zeng and Liang,2007).
Thus, the test statistics for Blindly Combined For the sample covariance matrix in equation
Energy Detection is given by; (14), let and be the maximum and
minimum eigenvalue of the covariance matrix,
if there is no
signal, . Thus, this
Where tr(.) is the trace of the matrix ratio can be used as the test statisticsto detect
and As shown from this equation, the
test-statistics is independent from the PU and noise the presence of signal (Zeng and Liang, 2007).
variances, as well as the channel gain vectors. Thus, In this regard, the test statistics for MME
this method is robust to synchronization error, detection is given by;
channel impairments and noise uncertainty
(Kortun,et.al,2010,Zeng and Liang,2008) Finally,
the algorithm for BCED is summarized as follows.
Algorithm 1: Blindly combined Energy Detection As it is shown, this detector also works without a-
1. Draw N-samples of the received signalX(N)= priori knowledge of the noise variance, thus it
[x(1),x(2),..., x(N)] fromM-Antennas. outperforms the multichannel energy detector under
2. Formulate and stack the ML-vector samples as noise uncertainty. This heuristic approach, however,
follows. does not fully exploit the inherent signal structure of
the signal covariance matrix and, consequently
degrades its detection performance (Zeng and
Liang,2007). The maximum-minimum eigenvalue
detection algorithm is given as follows.

Algorithm 2: Maximum-Minimum Eigenvalue


Detection
1. Draw N-samples of the received signal X(N)=
[x(1),x(2),..., x(N)]from M-antennas.
2. Formulate and stack the ML-vector samples
(same as step 2 in algorithm 1).
3. Compute the sample covariance matrix for this 3. Compute the sample covariance matrix for this
vector samples as : vector samples (use equation 14).

34
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
4. Sort and obtain the maximum and minimum i. The largest eigenvalue converges to;
eigenvalues of the covariance matrix.
5. Calculate the detection threshold from equation
(23)
6. Calculate the test statistics as : ii. The distribution of the largest eigenvalue
7. Decision: If TMME>τ, the primary user signal is converges to;
existed (decide in favour of H1), otherwise the
primary user is off.

A. Threshold Calculation for BCED and MMED


iii. The smallest eigenvalue converges to;
The threshold for eigenvalue based detection is
solved by using the concept of random matrix
theory (Chiani,2012). To do so, we first
introduce some useful results from Random iv. The distribution of the smallest eigenvalue
Matrix Theory (RMT) that express the limiting
distributions in which the largest and the converges to;
smallest eigenvalue of the sample covariance
matrix converge as the sample number (N), and
the product (ML) increases. Under H0, since the
columns of the received matrix are zero-mean
independent complex Gaussian vectors, the
sample covariance matrix is a complex Wishart Where TW2 is the Tracy Widom distribution
matrix (Chiani,2012). Through detail function of order 2 (Chiani, 2012). By using the
investigation of the fluctuations of the above asymptotic value and corresponding
eigenvalues of the Wishart matrices with RMT, distribution for the maximum and minimum
it can be understood that, when both the eigenvalues, we can calculate the detection
dimensions ofthe signal matrix tend to infinity threshold for MMED and BCED as follows.
with the same order, the eigenvalues of
matrices with random entries turn out to i. Detection Threshold for MMED
converge to some fixed distribution (Chiani, To find the detection threshold, let us use the test-
2012). Further, for Wishart matrices the statistics of the MMED detector from equation (15)
limiting joint eigenvalue distribution and the as;
marginal distributions of single ordered
eigenvalues have been found with RMT. Thus,
by exploiting some of these results, it can be
possible to express the asymptotical values of
the largest and the smallest eigenvalue of the
sample covariance matrix as well as their
limiting distributions (Chiani,2012) in the
following lemma that summarizes a number of
relevant results.

Lemma:Convergence of the smallest and largest


eigenvalues under : For M-antenna receiver
with oversampling factor (L) , and N time samples,
let us define the constant c as;

And assume that for ,


and let us define
(22)
(17) After substituting the values and rearranging,
we get;
When , the following results holds true.;
35
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

Rearranging terms, we get the threshold for BCED


as;

(23)
Where denotes the inverse of the
cumulative distribution function (CDF) of the
Tracy-Widom distribution of order 2. The
(28)
numerical table for the Tracy-Widom
distribution of order 2 is given below (Chiani,
2012) 3. The Spherical Test Detector (STD)
This detector is specifically useful for spectrum
Table 1: Numerical table for the Tracy-Widom sensing when the primary user covariance matrix is
distribution of order 2. unknown and when there are multiple primary
users. A GLRT procedure is applied to the problem,
x -1.80 -1.32 -0.59 -0.23 0.48 1.31
which finds unstructured estimate of the sample
F2(x) 0.50 0.70 0.90 0.95 0.99 0.999 covariance matrix (Rx), which is equivalent
to under H1,and under H0. The
ii. Detection Threshold for BCED resulting detector computes the Arithmetic to
Geometric Mean (AGM) of the eigenvalues of
The test statistics for the BCED detector in equation
sample covariance matrix and compares the result
(13) can be written in the form as ;
with pre-defined threshold. The test statistics for
spherical test detector is given by (Lim,et.al,2008);
where,

Now for large N, and ML , the value of The detection threshold is determined from a give
can be approximated by [19,20], thus ; probability of false alarm. Compared with the
multichannel energy detector with inaccurate
knowledge of the noise variance, the spherical test
detector is found to obtain improved spectrum-
sensing performance when the noise variance is
By using this, the test statistics in equation (24) can be unknown (Lim,et.al,2008).
approximated as 4. John’s Detector (JD)
A test statistics that is optimal for detecting small
deviations from H0 is John's detector. The criterion
under which John's detector is derived is known as
Under , the probability of false alarm for the the Locally Best Invariant (LBI) criterion
test in equation (24) is given by (Wie,et.al,2013). Unlike the GLR criterion, the LBI
criterion often leads to detectors that perform
particularly well in the low SNR regime. The
performance of this detector is also very good for
multiple primary user detection (Wie,et.al,2013).
The test statistics for John's detector is formulated
as;

IV. RESULTS AND ANALYSIS


In this section, we evaluate and analyze the
performance of the reviewed cooperative spectrum
sensing techniques. The analysis shows the impact
36
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
of the design parameters for the detection 400

algorithms. For simulationsetup, single primary


users, and multi antenna secondary users with 300

different number of receive antennas (M = 4, 6, 8)

PDF
are considered to evaluate the performances of 200

probability of detection and the probability of false-


100
alarms. The smoothing factor is taken as L=4, 8. A
BPSK modulated PU signal, and zero mean
0
complex Gaussian noise is assumed. A normalized 1 2 3 4 5
complex valued Rayleigh fading channel model Test(T)
with M-number of taps is considered. The Under H0
simulation results are averaged over 104 tests using a. Under H1

Monte-Carlo simulations written in Matlab.


500
Figure 3 shows the result of the detection
probability (Pd) versus signal-to-noise ratio (SNR) 400
for each detector. It can be shown from the figure
300
that for each detection algorithm the detection

PDF
probability increases with SNR. Covariance 200
Absolute Value based Detection (CAVD) is
included as reference. 100

0
0 5 10 15 20 25
Test(T)

Under H0
Under H1
b.
Fig. 4. Histograms of the test statistics for MMED with:

. For SNR=-10dB b). For SNR=0dB.

ROC Curve
1

0.8

Figure 3.Pd versus SNR for CAVD, MMED, BCED,


0.6
STD, JD with N=100, M=4,L=8, Pfa=0.1,MCR = 10 4,
in normalized complex Rayleigh fading channel.
Pd

Figure 4 shows empirical probability density n=40


0.4
functions (histograms) of the test statistic generated n=50
from the simulation for MMED with N = 100, M=4, n=60
L=8, and SNR = -10 dB and SNR = 0 dB. The 0.2
histograms plot hypotheses H0 (primary signal
absent) and H1 (primary signal present). It is noticed
that that at fixed threshold, Pd increases with SNR, 0
since the area of the histogram on the right of a 0 0.2 0.4 0.6 0.8 1
Pf a
given threshold under H1 increases. The net result is
an improved Pd for given Pfa.
Fig. 5.ROC curve for different number of samples with
Finally Figure 5 shows ROC curves for different M=4, L=8, SNR=-10 dB.
number of collected samples (N). It indicates the
relation between the probability of false alarm (Pfa)
and the probability of detection (Pd) for MMED. V. CONCLUSION
The figure shows that when the number of samples
increased, the detection performance of the In this paper, we reviewed robust multiple antenna
cognitive radio is improved. based spectrum sensing algorithms for CR networks
that works well under spatio-temporal correlation
between received samples. Based on the formulated
density functions, we obtained the sub-optimum
GLRT based detector for the case when the SU
37
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
receiver has no knowledge about the channel gains, 12. Zeng and Y. Liang(2007). Maximum-Minimum
the PU signal and noise correlation matrices. Then, Eigenvalue Detection for Cognitive Radio", Proc.
IEEE 18th Int. Symp. Personal, Indoor Mobile Radio
we formulated the test-statistics for eigenvalue Commun. (PIMRC), pp. 1-5.
based detections such as: BCED, MMED, STD, and
JD. Simulation results were provided to evaluate the 13. Nadler, F. Penna and R. Garello(2011). Performance
impact of the different parameters on the of Eigenvalue-based Signal Detectors with Known
performance of the stated detectors. The simulation and Unknown Noise Level. Proc. 2011 IEEE
results show that the performances of the reviewed International Conference Commun.
detectors are good and they are blind that work well
14. Huang and X. Huang(2013). Detection of
without a-priori knowledge of the PU signal, noise Temporally Correlated Signals over Multipath
variance and channel gains. Fading Channels", IEEE Trans. Wireless Commun.,
vol. 12, no. 3, pp. 1536-1276.
REFERENCES 15. Kortun, T. Ratnarajah, and M. Sellathurai(2010,Sep).
Exact Performance Analysis of Blindly Combined
1. Mitola& G. Q. Maguire (1999). Cognitive Radios: Energy Detection for Spectrum Sensing," in IEEE
Making Software Radios More Personal. IEEE PIMRC, (Istanbul, Turkey).
Personal Commun., vol. 6, no. 4, pp. 13-18.
16. Zeng and Y. C. Liang(2010). Robust Spectrum
2. Simon Haykin (2005,Feb). Cognitive Radio: Brain- Sensing in Cognitive Radio," in IEEE 21st
Empowered Wireless Communications. IEEE International Symposium on Personal, Indoor and
Journal on Selected Areas in Communications, vol. Mobile Radio Communications Workshops (PIMRC
23, no. 2, pp. 201-220. Workshops), pp. 1-8.

3. Yucek& H. Arslan (2009). A Survey of Spectrum 17. Alamgir , M. Faulkner , J. Gao and P. Conder(2008).
Sensing Algorithms for Cognitive Radio Signal Detection for Cognitive Radio Using Multiple
Applications. IEEE Communications Surveys and Antennas", Proc. IEEE Int. Symp. Wireless
Tutorials, vol. 11, no. 1, pp. 116–130, First Quarter. Commun. System.

4. Akyildiz , Brandon F. Lo &RavikumarBalakrishnan 18. Ramírez , J. Vía , I. Santamaría and L. L.


(2011,March). Cooperative Spectrum Sensing in Scharf(2010). Detection Of Spatially Correlated
Cognitive Radio Networks: A Survey. Physical Gaussian Time Series", IEEE Trans. Signal Process,
Communication, V.4 No.1, P.40-62. vol. 58, no. 10, pp. 5006-5015.

5. Digham, M.-S. Alouini, & M.K. Simon(2007). On 19. Chiani(2012). Distribution of the largest Eigenvalue
the Energy Detection of Unknown Signals Over for Real Wishart and Gaussian random Matrices and
Fading Channels. IEEE Transactions on a Simple approximation for the Tracy-Widom
Communications 55, 21–24. distribution. CoRR, vol. abs/1209.3394.

6. Zeng, Y.C. Liang, A. T. Hoang, and R. Zhang 20. Kay (1998). Fundamentals of Statistical Signal
(2010). A Review on Spectrum Sensing for Processing, Volume 2: Detection theory Prentice
Cognitive Radio: Challenges and Solutions. Hall PTR.
EURASIP Journal on Advances in Signal
Processing, vol., pp.1-15, January 2010. 21. Wei, P. Dharmawansa and O. Tirkkonen(2013,May).
Multiple Primary User Spectrum Sensing in the
7. Liang, Y. H. Zeng, E. Peh and A. T. Hoang (2008). Low SNR Regime. IEEE Trans. Commun., vol. 61,
Sensing–Throughput Trade-off for Cognitive Radio no. 5, pp. 1720-1731.
Networks. IEEE Trans. Wireless Communications.
vol. 7, no. 4, pp. 1326-1337.

8. Tandra and A. Sahai (2008). SNR Walls for Signal


Detection," IEEE J. Selected Topics Signal
Processing, vol. 2, no. 1, pp. 4-17.

9. Zeng and Y. C. Liang (2009). Spectrum-Sensing


Algorithms for Cognitive Radio Based On Statistical
Covariances. IEEE Trans. Veh. Technol., vol. 58, no.
4, pp. 1804-1815.

10. 10.Zeng , Y. Liang and R. Zhang (2008). Blindly


Combined Energy Detection for Spectrum Sensing
in Cognitive Radio. IEEE Signal Processing Letter,
vol. 15.

11. 11.Lim , R. Zhang , Y. C. Liang and Y. Zeng(2008).


GLRT-based Spectrum Sensing for Cognitive
Radio",Proc. IEEE Global Telecommun. Conf., pp.
1-5.
38
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

Joint Beamforming and Spectral Enhancement for Robust ASR


in Reverberant Environments
Fanuel Melak Asmare1, 2, Feifei Xiong1, Mathias Bode2, Bernd T. Meyer3 and Stefan Goetze1
1
Fraunhofer Institute for Digital Media Technology IDMT,
Group Hearing, Speech and Audio Technology (HSA), Oldenburg, Germany
2
Jacobs University Bremen, Bremen, Germany
3
University of Oldenburg, Dept. of Medical Physics and Acoustics, Oldenburg, Germany
melakeegzi@gmail.com, feifei.xiong@idmt.fraunhofer.de

Abstract — This work evaluates multi-microphone beamforming techniques and single-microphone


spectral enhancement strategies to alleviate the reverberation effect for robust automatic speech
recognition (ASR) systems in different reverberant environments characterized by different
reverberation times T60 and direct-to- reverberation ratios (DRRs). The systems under test consist of
minimum variance distortionless response (MVDR) beamformers in combination with minimum
mean square error (MMSE) estimators. For the later, reliable late reverberation spectral variance
(LRSV) estimation employing a generalized model of the room impulse response (RIR) is crucial.
Based on the generalized RIR model which separates the direct path from the remaining RIR, two
different frequency resolutions in the short time Fourier transform (STFT) domain are evaluated,
referred to as short-term and long-term, to effectively estimate the direct signal. Regarding to the
fusion between the MVDR beamformer and the MMSE estimator, the LRSV estimator can operate
either on the multi-channel observed speech signals or on the single-channel beamformer output. By
this, in this contribution four different combination system architectures are evaluated and analyzed
with a focus on optimal ASR performance w.r.t. word error rate (WER).
Key Words- Reverberation, Room impulse response, ASR

speech phonemes. Reduction of these detrimental


I. INTRODUCTION effects is evidently of considerable practical
importance. Although several pioneering efforts
Robust distant speech recognition greatly benefits were made (Kingsbury et.al. 1998; Seltzer, 2004),
those applications for which hands are not free to it is still very challenging to compensate for such
use and significantly increases the convenience of long-term distortions. Nowadays, improving the
use, e.g., the source can move freely while the robustness of the ASR systems in reverberant
microphone(s) is/are placed in a certain position in environments has been paid increased attention
distance to the speech source. Unfortunately, in (Yoshioka et al., 2012; Kinoshita et al., 2013;
distant talking scenarios the acoustic environment Xiong et al., 2014). Still, overcoming the
adds disturbances to the speech signal, i.e. problems reverberation problem is paramount for realizing
arise due to interfering noises as well as distant-talking speech recognizers.
reverberation caused by multiple reflections of the
desired speech signal at the room boundaries and The focus of this work is on removing the effects of
other objects in the room. This reverberation effect reverberation by pre-processing of the speech signal
degrades the speech quality and speech by a front-end processing procedure before the
intelligibility (Naylor, 2011), as well as deteriorates feature extraction phase of the ASR systems. By
the performance of automatic speech recognition applying different speech dereverberation strategies
(ASR) (Wölfel & McDonough, 2009). in single- and multi-microphone scenarios, we
intend to alleviate the reverberation effect in order
Reverberant speech can be perceived as sounding to improve the robustness of the ASR systems in
distant with noticeable coloration and echo. These different reverberant environments. We also use
detrimental perceptual effects generally increase different short time Fourier transform (STFT)
with increasing the room volume, as well as with lengths to analyze the separation of early reflection
increasing the distance between the source and the and late reverberation tail w.r.t. the performance of
microphone (Kuttruff, 2000). Furthermore, with the ASR systems.
spread in the time of arrival of reflections at the
microphone, reverberation causes blurring of

39
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
II. SYSTEM COMBINATION STRATEGIES III. SPECTRAL ENHANCEMENT FOR
DEREVERBERATION
As depicted in Fig.1, system (I) is based on a
minimum mean square error (MMSE) estimator as The observed reverberant speech signal can be
pre-processor in each of the input channels of the treated as the mixture of early reflections and the
beamformer. In this scenario, the late reverberation late reverberation, expressed in the STFT domain
spectral variance (LRSV) estimation is carried out as,
for the individual channels separately. Since the
pre-processor does not change the phase of the , (1)
signal, the spatial information for the beamformer where is the frame index and is the frequency
will not be affected. System (II) uses an bin. The early reflections are composed of
independent beamformer to get an enhanced the direct signal and early reflections, which are
spatially filtered signal followed by a single- usually set to 20-80 ms (Kuttruff, 2000). As well,
microphone dereverberation system. Since the early reflections play an important role for
LRSV and the MMSE estimators use this spatially enhancing the speech intelligibility. In contrast, late
filtered signal, the computational complexity is less. reverberation degrades the signal quality
The spatial filtering may, however, cause (Habets, 2007; Loizou, 2013). In this paper the
distortions to these estimator inputs due to the noise is neglected since the focus is to remove the
spatial correlation between microphone signals. To late reverberation. In order to suppress the late
avoid the spatial correlation effect on the LRSV reverberation effect, a spectral weighing function
estimators (Habets, 2007) in system (II), system determined by a parameterized MMSE
(III) uses a spatially averaged LRSV estimate estimator (Breithaupt et al., 2008) which shows
obtained from all the microphone signals refined by superior performance compared to the Wiener filter
the minimum variance distortionless response in our pilot experiments, is applied to the magnitude
(MVDR) beamformer. System (IV) illustrates a of the reverberant spectral variance in (1), resulting
LRSV in (Lebart, 2001),

MMSE (2)
MVDR LRSV
beamformer where is a lower bound of the weighting
MMSE
MVDR function. The a-priori early reflection to late
beamformer MMSE
reverberation energy rationecessary in (2) to
LRSV (I) (II) calculate is estimated by the decision-
directed (dd) approach (Ephraim & Malah, 1984)
(III) (IV) which performs slightly better than the temporal
MVDR MVDR cepstrum smoothing technique used in (Xiong et al.,
beamformer MMSE MMSE
beamformer
2014) for dereverberation in our pilot experiments.
Post-Filtering By this the LRSV estimation is required for
LRSV (2) as also shown in Fig. 1. A generalized statistical
d LRSV reverberation model (Habets et al., 2009) is used
here which separates the direct path from Polack’s
room impulse response (RIR) model (Lebart, 2001),
Fig. 1: Four different system combinations (I)-(IV) resulting in the spectral variance of the RIR in
consisting of the MVDR beamformers and the MMSE the STFT domain as
estimators, as well as the respective LRSV estimators.
multi-channel MMSE enhancement scheme (3)
(Simmer et al., 2001), which can be decomposed
into an MVDR beamformer followed by a single- where the decay coefficient is related to the
channel Wiener filter. Here the MMSE estimator is reverberation time by . and
actually a post-filter (. McCowan & Bourlard, 2003; denote the variances of the direct path and the
Zelinski,1988). reverberant part and represents the STFT frame
shift, i.e. the hop size. The direct signal to
reverberation ratio ( ) can then be expressed as
(Habets et al., 2009)
40
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
Simmer, 2001) or by the identity matrix for the
(4) delay and sum (DS) beamformer. A post-filtering
approach is used to calculate the coefficients in
The is related to the clarity index due to the system (IV) (McCowan & Bourlard, 2003) which
frame shift (Kuttruff, 2000). Now the can be expressed as an MVDR beamformer with a
reverberation variance can be obtained using (4) as post-filter with the post filter
(Ephraim & Malah, 1984) (ignoring the frequency transfer function (Zelinski,1988) expressed by
index for simplicity)

(5)
(9)
where is calculated from the in (4),
constraint inthe range of (0, 1]. Then, the LRSV is with being the auto-correlation of the speech
given by signal in microphone channel . In order to alleviate
speech distortions in ASR systems, a lower bound
, (6) is introduced. The early reverberation variance
where denotes the number of frames which (the cross correlation term) can be estimated as
corresponds to the duration of early reflections of (McCowan & Bourlard, 2003)
the RIR. An instantaneous estimate of the input
reverberant spectral variance in (5) can be
obtained by a smoothed version of as
, (7)
(10)

where the smoothing constant is calculated by where calculates the real part of a complex
. According to (Habets, 2007), in signal. A time alignment is required for when
order to improve the tracking performance of the calculating (10) (McCowan & Bourlard, 2003),
reverberant speech onset, shall be set to be lower which can be achieved by the steering vector as
seen in Fig. 1 (IV). In (10) a first-order recursive
than when . Note that such an
LRSV estimator requires a-priori information of update of the auto- and cross-correlation
and DRR or clarity index at least in full-bandmode, calculations is applied and a maximal threshold is
which in practice can be estimated by (Eaton et al., introduced to avoid the denominator being non-
2013; Xiong et al., 2013). positive. The LRSV coherence matrix in (8) can
also be replaced by the LRSV coherence
estimated from the received microphone signals.
IV. DEREVERBERATION BY MULTI-
MICROPHONE BEAMFORMING AND POST-
FILTERING V. EXPERIMENTAL RESULTS

When multiple microphones are available, The WSJCAM0 British English corpus (Robinson
beamforming and post-filtering techniques can be et al., 1995) was used as database of clean
used for the purpose of dereverberation (Allen et. (anechoic) speech utterances. It contains 7861
al. 1977; Marro et al., 1998; Jeub & Vary, 2010; utterances for training and another 742 for testing at
Westermann et al., 2013). The MVDR beamformer a sampling rate of 16 kHz. 18 real-world RIRs
which performs best in diffuse interference fields is recorded by a circular microphone array (M = 8)
used here. This beamformer minimizes the output with 20 cm diameter from the REVERB Challenge
power while keeping a unity gain in the desired (Kinoshita et al., 2013) were used for multi-
direction and its coefficients can be derived as, condition training mode and another 6 RIRs
(Kinoshita et al., 2013) for generating various test
, (8) sets (denoted by T1-T6 in the following) with
where is the Hermitian transpose and is the different T60 and DRR values, as listed in Tab. 1.
steering vector. In order to vary the beamformer The STFT was computed using a Hanning window
used, the coherence matrix of the interfering signals with two different analysis window lengths, 32 ms
is replaced by a diffuse interference field for with 1/8 overlap (short term) and 96 ms with 1/2
the superdirective (SD) beamformer (Bitzer & overlap (long term). A white noise gain constraint

41
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
of 10 dB was selected for the MVDR beamformer improve the ASR output, with the SD beamformer
in (8). The weighting factor of 0.5 was used in the showing lower WERs of approximately 4%
dd approach. was chosen as in (7). in compared to the single microphone MMSE
(2) was set to -10 dB as a good value to the ASR estimator.
performance. in (9) was chosen as 0.1 and the Table 1. Characteristics of all test sets T1-T6 with mean
smoothing factor of the first-order recursive filter DRR and T60 values (from all 8 channels).
was 0.875 in (10). Directly from the RIRs in full-
band mode, DRR or was calculated
Test T1 T2 T3 T4 T5 T6
accordingly and was determined by using
Schroeder’s method (Schroeder, 1965). Room Small Small Mediu Mediu Large Large
m m
The framework for the ASR experiments was
Positio
implemented based on the Hidden Markov Model n Near Far Near Far Nea Far
Toolkit (HTK) (Young et al., 2009). Overlapping DRR
speech segments of 25 ms duration and 10 ms shift (dB) 17.73 4.56 11.42 0.25 10.40 -1.70
were used for the calculation of mel-frequency T60
cepstral coefficients with delta and double-delta (ms) 218.29 229.9 500.06 519.44 719.3 747.3
coefficients as well as cepstral mean and variance 1 2 8

normalization. Context-dependent triphone hidden The performances of system (I) and the MVDR-
Markov models with 3 states per model were only system show very similar results. For test sets
applied together with 12 Gaussian mixture models T1, T3 and T5, i.e. the near-position tests, higher
per state and a language scaling factor of 14.0 for WERs can be observed when compared to the SD
the 5k-word-bigram language model. beamformer alone. This may be caused by the
Fig. 2 shows the word error rate (WER) results of distortions of the diffuse field because of the front
our systems under test with 32 ms STFT analysis MMSE estimators. Compared to system (I), 1%
window length. The baseline results come from the WER improvement can be obtained by system (II).
multi-condition training with the original A more accurate LRSV estimate is employed in
reverberant speech signal from the first microphone, system (III) which results in a slightly better
i.e. i=1. For single-microphone scenarios, the performance than system (II). This indicates that,
MMSE estimator is applied to the first channel. For the spatial correlation introduced by the
comparison, the results with beamformers alone beamformer blurs the MVDR-filtered RIR in
(either with SD or DS) are presented. The rest are system (II) so that it cannot exactly extract the true
the system combination outputs (cf. Fig. 1). Both late reverberation.
the MMSE and the beamformer alone scenarios

35 baseline (reverb)
MMSE only
DS only
30 MVDR only
(I)
(II)
25
WER %

(III)
(IV) with MVDR
(IV) with LRSV
20

15

10
T1 T2 T3 T4 T5 T6
Test Sets

Fig.2 WERs of the dereverberation strategies with 32 ms STFT analysis window length and 1/8 hop size.

42
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
26
(II) 32 ms 1/8 hop
24
(II) 96 ms 1/2 hop

22 (III) 32 ms 1/8 hop

(III) 96 ms 1/2 hop


20
WER %

18

16

14

12

T1 T2 T3 T4 T5 T6
Test Sets

Fig.3 WERs of system (II) and (III) with two different frame lengths, i.e. short and long term STFT analysis window
length. Male, Far
4
reverb
MMSE
MVDR
3.5
(I)
(II)
(III)
3
PESQ

(IV) (MVDR)
(IV) (LRSV)

2.5

T1 T2 T3 T4 T5 T6
Test Sets

Fig.4 PESQ scores from the output of different systems; a male utterance from the test data is employed and the
respective clean (anechoic) speech is used as the reference signal.

Overall, average WER reduction of 6.17% is that the 32 ms with 1/8 hop window improves the
obtained by system (III) with the SD beamformer WER by 4.18% and 3.57% compared to the 96 ms
compared to the baseline. Such improvements window with 1/2 overlap window with systems (II)
become more obvious for the far-position testsets and (III) respectively.
such as T4 and T6 than the near-position test sets
A perceptual evaluation of speech quality (PESQ)
such as T3. A similar trend can be observed for
(ITU-T, 2001) has also been conducted. Fig. 4
system (IV), for which the SD beamformer still
shows the PESQ scores of the different proposed
performs best. It can also be observed that the
systems with one male test utterance. The
results degrade when the beamformer in (1) uses the
performance of multi-microphone dereverberation
LRSV coherence matrix. A possible explanation is
strategies in PESQ tests is much better than that of
that the late reverberation behaves non-stationary
single-microphone approaches. Here system (III)
and its coherence actually does not match the
shows the best results compared to all other
diffuse property, especially for the near-position test
scenarios, which is in consilience with the WERs
sets T1, T3 and T5 as discussed in (Xiong et al.,
results in Fig. 2.
2015).
Fig. 3 compares the WERs of the two best VI. CONCLUSION
performing systems (II) and (III) with two different
STFT analysis window sizes. Using shorter frame This work explored possible combination
sizes benefits in reducing WERs. The results shows architectures for dereverberation by (single-
microphone) spectral enhancement schemes and
43
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
(multi-microphone) beamforming with the aim of Filterbank for Feature Extraction,”REVERB
improving ASR performance in various reverberant challenge, Florence, 2014.
environments. Results indicate that all the combined 9. E.A.P. Habets, “Single- and Multi-Microphone
systems are able to provide benefits for ASR Speech Dereverberation using Spectral
systems and specifically, the system (III) combining Enhancement,”Ph.D. thesis, University of
the SD beamformer and the MMSE estimator with Eindhoven, Eindhoven, The Netherlands, Jun. 2007.
the LRSV refinement by the MVDR beamformer
coefficients achieves nearly 30% average relative 10. K.U. Simmer, J. Bitzer, and C. Marro, “Microphone
Arrays”, Chapter “Post-Filtering Techniques,” pp.
WER improvement compared to the baseline, as
39–60, M. Brandstein and D. Ward (Eds.), Springer,
well as 15% average relative PESQ boost (from one
Berlin, Heidelberg, May 2001.
example) compared to the first channel reverberant
speech signal. As well, short STFT analysis window 11. I.A. McCowan and H. Bourlard, “Microphone Array
length provides better ASR performance than a Post-Filter based on Noise Field Coherence,” IEEE
longer window length. Transactions on Speech and Audio Processing, vol.
11, no. 6, pp. 709–715, Nov. 2003.

LITERATURE 12. R. Zelinski, “A Microphone Array with Adaptive


Post-Filtering for Noise Reduction in Reverberant
1. P.A. Naylor, N.D. Gaubitch, and E. Cross, “Speech Rooms,” in IEEE Int. Conf. on Acoustics,Speech,
Dereverberation,”Noise Control Engineering and Signal Processing (ICASSP), New York, NY,
Journal, 59(2): 211–211, 2011. USA,Apr. 1988, vol. 5, pp. 2578–2581.
2. M. Wölfel and J. McDonough,“Distant Speech 13. P.C. Loizou,“Speech Enhancement: Theory and
Recognition,”John Wiley & Sons Ltd, 2009. Practice,”CRC press, 2013.
3. H. Kuttruff, “Room Acoustics,”Spon Press, London, 14. C. Breithaupt, M. Krawczyk, and R. Martin,
4th edition, 2000. “Parametrized MMSE Spectral Magnitude
4. B.E.D. Kingsbury, N. Morgan, and S. Estimation for the Enhancement of Noisy Speech,”
Greenberg,“Robust Speech Recognition using the in IEEE Int. Conf. on Acoustics, and signal
Modulation Spectrogram,”Speech communication, Processing (ICASSP), Las Vegas, NV, USA, Apr.
25(1):117-132, 1998. 2008, pp. 4037-4040.

5. M.L. Seltzer, B. Raj, and R.M. Stern,“Likelihood- 15. K. Lebart, J.M. Boucher, and P.N. Denbigh,“A New
Maximizing Beamforming for Robust Hands-Free Method based on Spectral Subtraction for
Speech Recognition,”Speech and Audio Processing, Speech,”ActaAcusticaunited with Acustica,
IEEE Transactions on, 12(5):489–498, 2004. 87(3):359–366, 2001.

6. T. Yoshioka, A. Sehr, M. Delcroix, K. Kinoshita, R. 16. Y. Ephraim and D. Malah, “Speech Enhancement
Maas, T. Nakatani, and W. Kellermann,“Making using a Minimum Mean-Square Error Short-Time
Machines Understand Us in Reverberant Rooms: Spectral Amplitude Estimator,” IEEE Transactions
Robustness Against Reverberationfor Automatic on Acoustics Speech and Signal Processing, vol. 32,
Speech Recognition,”IEEE Signal Processing no. 6, pp. 1109–1121, Dec. 1984.
Magazine, 29(6):114–126,2012. 17. E.A.P. Habets, S. Gannot, and I. Cohen, “Late
7. K. Kinoshita, M. Delcroix, T. Yoshioka, T. Reverberant Spectral Variance Estimation based on
Nakatani, E. Habets, R. Haeb-Umbach, V. Leutnant, a Statistical Model,” IEEE Signal Processing
A. Sehr, W. Kellermann, R. Maas, S. Gannot, and B. Letters, vol. 16, no. 9, pp. 770–773, Sep. 2009.
Raj,“The Reverb Challenge: A common Evaluation 18. J. Eaton, N.D. Gaubitch, and P.A. Naylor, “Noise-
Framework for Dereverberation and Recognition of Robust Reverberation Time Estimation using
Reverberant Speech. IEEE Workshop on Spectral Decay Distributions with Reduced
Applications of Signal Processing to Audio and Computational Cost,” in IEEE Int. Conf. on
Acoustics (WASPAA), 2013. Acoustics, Speech, and Signal Processing (ICASSP),
8. F. Xiong, N. Moritz, R. Rehr, J. Anemüller, B.T. Vancouver, Canada, May 2013, pp. 161–165.
Meyer, T. Gerkmann, S. Doclo, and S. 19. F. Xiong, S. Goetze, and B.T. Meyer, “Blind
Goetze,“Robust ASR in Reverberant Environments Estimation of Reverberation Time based on Spectro-
using Temporal Cepstrum Smoothing for Speech Temporal Modulation Filtering,” in IEEE Int. Conf.
Enhancement and an Amplitude Modulation on Acoustics, Speech, and Signal Processing

44
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017
(ICASSP), Vancouver, Canada, May 2013, pp. 443– D. Povey, V. Valtchev, and P. Woodland, “The HTK
447. Book (for HTK Version 3.4),”Cambridge University
Engineering Department, Cambridge, 2009,
20. J.B. Allen, D.A. Berkley, and J. Blauert, “Multi-
http://htk.eng.cam.ac.uk/.
microphone Signal Processing Technique to Remove
Room Reverberation from Speech Signals,” J. 28. F.Xiong, B.T. Meyer, and S. Goetze, “ A Study on
Acoust. Soc. Am., vol. 62, no. 4, pp. 912–915, 1977. joint Beamforming and Spectral Enhancement for
Robust Speech Recognition in Reverberant
21. C. Marro, Y. Mahieux, and K.U. Simmer, “Analysis
Environments,” in IEEE Int.Conf. on Acoustics,
of Noise Reduction and Dereverberation Techniques
Speech, and Signal Processing (ICASSP), Brisbane,
based onMicrophone Arrays with Postfiltering,”
Australia, Apr. 2015.
IEEE Trans. Speech and Audio Processing, vol. 6,
no. 3, pp. 240–259, May 1998. 29. ITU-T, “Perceptual Evaluation of Speech Quality
(PESQ): An Objective Method for End-to-End
22. M. Jeub and P. Vary, “Binaural Dereverberation
Speech Quality Assessment of Narrow-Band
based on A Dual-Channel Wiener Filter with
Telephone Networks and Speech Codecs,” Feb.
Optimized Noise Field Coherence,” in IEEE Int.
2001.
Conf. on Acoustics, Speech, and Signal Processing
(ICASSP), Dallas, TX, USA, Mar. 2010, pp. 4710–
4713.
23. A.Westermann, J.M. Buchholz, and T. Dau,
“Binaural Dereverberation based on Interaural
Coherence Histograms,” J. Acoust. Soc. Am., vol.
133, no. 5, pp. 2767–2777, 2013.
24. J. Bitzer and K.U. Simmer, “Microphone Arrays,
chapter Superdirective Microphone Arrays,” pp.
19–38, M. Brandstein and D. Ward (Eds.), Springer,
Berlin, Heidelberg, May 2001.
25. T. Robinson, J. Fransen, D. Pye, J. Foote, and S.
Renals, “WSJCAM0: A British English Speech
Corpus for Large Vocabulary Continuous Speech
Recognition,” in IEEE Int. Conf. on Acoustics,
Speech, and Signal Processing (ICASSP), Detroit,
Michigan, USA, May 1995, pp. 81–84.
26. M.R. Schroeder, “New Method of Measuring
Reverberation Time,” J. Acoust.Soc. Amer., vol. 37,
no. 3, pp. 409–412, 1965.
27. S. Young, G. Evermann, M. Gales, T. Hain, D.
Kershaw, X.A. Liu, G. Moore, J. Odell, D. Ollason,

45
Joint Beamforming and Spectral Enhancement for Robust ASR in Reverberant Environments
ICAST-2017

46
Modeling and Analysis of Linear Observer Based On DC Machine
ICAST-2017

Modeling and Analysis of Linear Observer Based On DC Machine


Debabrata Pal1, Ejigu Tolina2
1
Department Of Electrical & Computer Engineering, Axum, 1010,Ethiopia,Email: Debuoisi@Gmail.Com
2
Department Of Electrical & Computer Engineering, Axum, 1010,Ethiopia,Email: Ejigutolina@Gmail.Com

Abstract — This paper presents how to implement full-order and reduced-order linear observers
using the software packages for computer aided control system design MATLAB. Observers are used
these days not only for the purpose of feedback control, but also in their own right to observe state
variables of a dynamic system, which can be an experiment in progress whose state has to be
monitored at all times. As a matter of fact, it has shown how to implement a dc machine state space
model and its observer. Different observer gain(s) are determined by selecting different eigen values
for the observers. Observer estimation errors are presented by choosing the observer(s) initial
conditions. Having full understanding of the observer implementation(s), students and engineers will
fill confident to use this observers and observer based controllers in numerous engineering and
scientific applications.
Key words- Full Order Observer, DC Machine, Reduced Order Observer, MATLAB/SIMULINK
form, where the state functional are treated as
additional outputs. (Tsui, 1985) developed an
I. INTRODUCTION
interesting algorithm for the design of
Estimating linear functional of state vector has been multifunctional observer. The same author (1986)
the focus of many researchers over the years. A showed that the order of observer can be reduced to
number of procedures have been proposed to design the sum of descending ordered observability indices
linear functional state observer. (Fairman and of the system minus the number of the required
Gupta, 1980; O’Reilly, 1983; Aldeen and Trinathh, functional. (O’Reilly, 1983) provides excellent
1999; and Trinath and Ha, 2000). It is often notes on observer theory in chapter 3. In this paper,
convenient when designing feedback control system a new observer capable of asymptotically estimating
to assume initially that the entire state vector of the any vector state functional, is introduced the order
system to be controlled or it is available for of the observer is dependent on the ratio of the
feedback. number of independent output measurements to the
number of independent outputs. Same method we
If the entire state vector cannot be measured, in
can extend for time delay and system with unknown
most complex systems, the control law cannot be
inputs.
implemented. In such a case either a new approach
that directly accounts for the non availability of the The rest of the paper is organized as follows: The
entire state vector must be devised, or a suitable modeling and analysis of Full order (II) and
approximation to the state vector must be Reduced Order (III) Observer is shown.DC machine
determined. In almost every situation the latter state model is derived in section (IV) and parameter
approach, that of developing and using an values are assigned (V).MATLAB program is used
approximate state vector is vastly simpler than a for determination of observer gain (VI).Result and
new direct attack on the design problem. The discussion is also shown in next section(VII).Finally
possibility of constructing an observer was first a conclusion is given at section (VIII).
explored by Bass and Gura (1965) and Luenberger
(1964,1966,1971).Alterative design procedures for
II. FULL ORDER OBSERVER MODELING
observing a scalar linear function of the state of a
multiple output system are minimal partial The theory of observers originated in the work of
realization theory and decision method respectively. Luenberger in the middle of 1960s.According to
Design procedures are suggested for observing a Luenberger, any system driven by the output of the
scalar linear function of a state of a multiple output given system can serve as an observer for that
system by (Moore, 1975), (Murdoch, 1973), and system. Consider a linear dynamic system with
(Roman, 1973). (Fortmann and Williamson, 1972) unknown initial value of
are the first to reconstruct vector linear functionals
= , (1)
of the state for multiple output systems. Another
method based on the reduction of a state observer A full order state observer estimates all of the
for a system in form, where the states functional are system state variables. In practice this may not
treated as additional outputs, was proposed by happen for a number of reasons including cost or
(Fairmann and Gupta,1980) Fairmann and Gupta that the state may not physically be measurable.
suggest another method based on reduction of state
observer for a system in Luenberger companion

47
Modeling and Analysis of Linear Observer Based On DC Machine
ICAST-2017
We define the mathematical model of the observer =
to be
= + (6)
= + (2)
Where = = estimation error
Where is the estimated state and C is the
estimated output. The n×1 matrix Ke is called state Combining equations (3) and (6), we obtain
observer gain matrix. Hence the observer error
equation is defined by = (7)
- = ( - ) (3)
The above equation describes the dynamic of the
= e observed state feedback control.
Where ( - ) = e = observer estimation error
vector. Thus the dynamic behavior of the error III. REDUCED ORDER OBSERVER MODELING
vector depends upon the Eigen values of
. A reduced-order state observer estimates fewer than
If the observer gain Ke is chosen such that the n state variables where n is the dimension of the
feedback matrix is asymptotically stable state vector. Consider equation (1) where the state
(has all eigen values with negative real parts) then vector
the estimation error e will decay to zero for any can be partitioned into two parts (a scalar) and (a
initial condition of e. This stabilization requirement vector). Here the state variable is equal to the
can be achieved if pair (A,C) is observable. The output y and thus can be directly measured and is
observer eigen values should be chosen to be about the unmeasurable portion of the state vector. Then
ten times faster than the system eigen values. This partitioned state and output equations become
can be achieved by setting the smallest real part of
the observer eigen values to be ten times bigger than = + u (8)
the largest real part of the closed loop eigen values
> y= (9)

10 (4) From equation (8) the equation for the measured


portion and unmeasured portion of the state
Theoretically, an observer can be made arbitrarily
becomes.
fast by
pushing its closed loop eigen values very far to the = + + u (10)
left in the complex plane , but very fast observers
generate noise, and this is not desirable. In control = + + u (11)
system practice, instead of ten times, it is sufficient Here equation (10) and (10) are known as “state
that the closed loop observer eigen values are faster equation” and “output equation” for the minimum
than the closed loop system eigen values 5-6times. order observer.
As we know the equation for the full order observer
It is important to point out that the system observer is
configuration preserves the closed loop system
eigen values that would have been obtained if the = + (12)
linear perfect state feedback control had been used. Then making the substitution of table 1 into last equation
This fact will be shown in the follow up section. we obtain:
The system (1) under the perfect state feedback
=( - ) + + +
control, that is has the closed loop form
(13)
= (5)
So the eigen values of the matrix are the By subtracting equation (13) from equation (11) we
closed loop eigen values under perfect state obtain
feedback. - =( - )( - ) (14)
For state feedback control , it is known that
Define e = -

Then equation (13) becomes


=( - )e (15)
With this control, the state equation (13) becomes
=
48
Modeling and Analysis of Linear Observer Based On DC Machine
ICAST-2017
+ = (20)
Table1: List of necessary substitutions for writing the
observer equation for the minimum order state observer Where = armature voltage, = back emf,
= armature current, = developed
Full Order State Reduced Order State
Observer Observer torque, ) = motor shaft angle, = =
shaft speed, = moment of inertia of the rotor, B=
A viscous friction, = inductance of armature
Bu + U windings, = armature winding resistance, =
y - + - motor torque constant, = motor constant.
u
C Linear observer is designed based on DC machine
(n×1) matrix [(n-1)×1] (motor) state model. Angular speed of the motor
and armature current are taken as state variable.
Thus the dynamic behavior of the error vector Therefore = and
depends upon the Eigen values of -
This is the error equation for the minimum order =
observer. The dynamic behavior of error vector
depends upon the Eigen values of - . State model of the dc machine (motor) is given
Now the characteristic equation for the minimum below:
order observer is obtained the equation as follows.
=(s- )(s- )………..(s- ) = + u (t)
=0
Where, are desired Eigen
values for the minimum order observer. Suppose the = (21)
desired location of the Eigen value for the minimum
observer is at s = - 9. The system dynamics for
reduced-order observer is defined by the given V. STATE MODEL OF MACHINE USING
equation. PARAMETERS

= (16)
Let, the motor parameters (coefficient of differential
equations) are assigned to be = 0.5 H, = 0.01
N-m/A, = 0.01 V-sec/rad, J = 0.01 kg-m2, B =
IV. STATE MODELING OF DC MACHINE 0.1 N-m-sec/rad, = 1 Ω.
Direct current machines are the most versatile
Thus the state model of dc machine is derived using
energy conversion devices. Their outstanding
advantage is that the volt-ampere or speed torque machine parameters and equation (21) as follows:
characteristic of these machines are very much
flexible and easily adaptable for both steady state = + u (t)
and dynamic operations. When a wide range of
speed control and torque output are required dc
motor is an obvious choice. = (22)

The state space approach is a generalized time


domain method for modeling, analyzing and VI. MATLAB PROGRAM FOR OBSERVER
designing a wide range of control systems and is GAIN ASSIGNMENT
particularly well suited to digital computational
technique. In this paper armature current and speed Ackerman‟s formula is used to write the program in
of the dc motor are taken as state variables. MATLAB command window. The programs are
The different equations related to DC motor are given below.
given below:
>> %State feedback matrix K design using pole-
= (17) placement technique
= + + (18) A = [-10 1;-0.02 -2];
= (19) B = [0;2]; C = [1 0]; D = [0];

49
Modeling and Analysis of Linear Observer Based On DC Machine
ICAST-2017
J = [-2+j*1 -2-j*1];
K = acker(A,B,J) It can be seen from Figures 1 and 2 that the reduced-
order observer is superior over the full-order
K = 32.4900 -4.0000 observer because the reduced-order observer is
more accurate than the corresponding ones observed
by the full-order observer. Note that the eigen
>> %State observer gain matrix Ke design (Full values for both observers are placed to be of the
Order) same speed. The reduced- order observer value is
placed at -9 and the full-order observer values are
>> A = [-10 1;-0.02 -2]; placed at -9,-10. Even more, the reduced-order is
B = [0;2]; C = [1 0]; D = [0]; simpler for implementation since they are
L= [-9 -10]; dynamical systems of lower order that the original
Ke = acker (A',C',L)' systems.

Ke =
VIII. CONCLUSION
7.0000
55.9800 It has shown in detail how to implement full-order
and reduced-order observers in MATLAB
environment and presented corresponding
>> % State observer gain matrix Ke design fundamental derivation and results with the help of
(Reduced Order) dc motor state model. I hope this paper will
>> Aab = [1]; Abb = [-2]; motivate undergraduate and graduate students to
>> LL = [-9]; further study observers and use them as powerful
>> Ke = acker (Abb', Aab', LL)' tools for observing system dynamics and/or
designing feedback control loops.
Ke = 7

REFERENCES
VII. RESULTS AND DISCUSSION
1. Bass, R. W., and Gura (1965). ‘High-order system
The obtained differences between the actual state design via Automatic Control Conference, Atlanta,
trajectories and estimated state trajectories using the Georgia. Luenberger, D.G. (1971). ’An introduction
full and reduced-order observers for dc machine are to observer’,IEEE Tran Auto. Control, AC-16, pp.
presented in Figures 1 and 2 using MATLAB 596-602
environment. In both cases the initial conditions are
taken as ; x1(0) = 1, x2(0) = 0, e1 (0) =1, e2(0) = 2. Luenberger(1966).’Observers for multi-variable
systems’, IEEETranAuto.Control, AC-11, pp.190-
0,e(0) = 1.
197

1
3. Moore, J.B. and Ledwitch G.F(1975).’Minimal order
observers for estimating linear functionals of state
vector’, IEEE TranAuto. Control, AC-20,p p. 623-
0.5
e1

0 632.
-0.5
0 1 2 3 4 5 6 7 8 4. Murdoch, P.(1973).’Observer design of a linear
functional of the state vector’, IEEE Tran
Time(sec)

0
Auto.Control, AC-18,pp.308-310 Roman, J.R.,
-1
Jones, L.E., and Bullock,T.E,(1973). ‘Observing a
e2

-2 function of the state’, Proceedings of IEEE Decision


-3
and Control Conference, San DiegoCalifornia,
0 1 2 3 4 5 6 7 8
Time(sec)
5. Fortman, T.E., and Williamson D (1972). ’Design of
Fig.1. Full-order observation estimation error low order observer for linear feedback control
laws’, IEEE Tran Auto.Control, AC-18, pp.301-308.
1

0.8 6. Fairman F.W.and Gupta R.D.(1980).’Design of


0.6
multifunctional reduced order
e

0.4

0.2 observer’.Int.J.Sys.Sci. 11, pp.1083-1094.


0
0 1 2 3 4 5 6 7 8
Time(sec)

Fig.2. Reduced -order observation estimation error

50
Modeling and Analysis of Linear Observer Based On DC Machine
ICAST-2017
7. Tsui C.C.(1985)’A new algorithm for the design of
multifunctional observers’, IEEE Tran Auto.Control,
AC- 30,pp.89-93.

8. Tsui (1986), C.C.,’on the order reduction of linear


functional observers’, IEEE TranAuto.Control, AC-
31, pp.447-449. O’Reilly J.(1983)’Observe for linear
systems’. Academic Press,
9. Aldeen(1999)..’Reduced order linear functional
observer for linear system’.IEEE Proc.Control
Theory Appl., Vol.146, No.5.
10. H.Trinath and Q.Ha(2000)’Design of linear
functional observers for linear systems with
unknown inputs’. International Journal of System
Sciene.’Vol 31, pp.741-749.

51
Modeling and Analysis of Linear Observer Based On DC Machine
ICAST-2017

52
Distributed Adaptive Cruise Control
ICAST-2017

Distributed Adaptive Cruise Control


Solomon Genene Gudeta1
1
Department of Electromechanical Engineering, Addis Ababa Science and Technology University, Addis Ababa,16417,
Ethiopia, bbgudeta@gmail.com

Abstract — Adaptive cruise control (ACC) is crucial part of vehicle longitudinal control in driver assistance
system (DAS). It has adaptive cruise states, engine control unit, brake control unit and drive train unit as sub
parts with different sensors and actuators connected to them. There are low level controllers in engine
control and brake control units to make the car adapt to traffic environment. The FMCW radar is used to
detect relative speed and relative range of target car in front of ACC car. The designed adaptive cruise
control components communicate over controller area network (CAN) making a distributed control system
(DCS). Ultimately, the designed distributed control system is implemented using simulink and simscape
blocks. Unlike the conventional adaptive cruise control, the designed ACC can be used in a low speed
scenario and it is verified using simulation results.
Key words- Adaptive Cruise Control, Controller area network, throttle control, Anti-lock Braking
System, vehicle dynamics
conventional braking system. ABS system brakes
I. INTRODUCTION
Futile road accidents that can be avoided by alerting
drivers to the situations are so many in numbers.
There are different driver assistance systems that
informs driver about the in front road information
and thereby assisting drivers in avoiding/reducing
accidents that could come along with it (F.
A.Arvind Raj R et al, 2011). Adaptive cruise control
(ACC) and anti-lock braking systems (ABS) are few
of such a driver assistances systems (DAS).
Adaptive cruise control system adapts the vehicle's
speed to the traffic environment while making the Fig.1. ACC Vehicle relationships
car to move at the desired speed or desired time gap
to the target vehicle. It makes the car to move with a the car without locking the wheel which results in
speed specified by the driver if there is no other small braking distance as the car doesn't supposedly
vehicle in the car's lane; and it also maintains the slide.So, the car can be handled and maneuvered by
minimum time gap to the front car if the front car is the driver as desired. If deceleration of the ACC car
in the same lane. The minimum time gap and the is not sufficient enough to avoid a collision, the
desired cruise speed are set by the driver. The ACC warns the driver with a warning sound. If the
conventional ACC is not used in low speed system detects the target vehicle is no longer in the
scenarios. The car with ACC is called ACC car ACC car's path, the ACC will accelerate the car to
while the in front vehicle in the lane of ACC car is desired speed set by the driver. So, this operation
called target vehicle. An anti-lock brake system makes the car to slow down and speed up with the
(ABS) controls the slip of each wheel to prevent it traffic. The ACC car speed is controlled through
from locking as a result of the service brake being engine throttle control and limited brake operation.
applied with too much force (Idar et al, 2001). The The ACC was often used in free driving with speed
idea is to maintain cornering forces on braked between and (F. A.Arvind
wheels to ensure that the vehicle retains its driving Raj R et al, 2011). It was never used in slow driving
stability and maneuverability as far as physically with stop and go scenario. The normal ACC doesn't
possible. work under the speed below 30 to 50 (F.
In order to maintain the time gap to the target A.Arvind Raj R et al, 2011). The cruise control
vehicle, ACC uses radar system and ABS system. operation using radar sensors is one of the sensing
Radar system is used to detect the relative speed of technologies used for automotive control (Camilla
target vehicle and range to target vehicle. If target et al, 2009). Radar sensors are so important that
vehicle is moving slowly, the ACC system will slow they perform equally well during the day, the night,
down the ACC car and controls the time gap and in most weather conditions. Though passive
between the ACC car and target vehicle. While optical devices can be an alternative to radar
slowing down the vehicle, the ACC system applies sensors, they have less accuracy at greater distances.
brake using ABS system rather than the

53
Distributed Adaptive Cruise Control
ICAST-2017

In this paper distributed ACC is designed Where and are longitudinal tire forces at
considering low speed scenario as low as 6 . front and rear wheels respectively, is an
The ACC module, the brake control unit, the engine aerodynamic drag force, is mass of the vehicle,
control unit are designed separately. High speed is longitudinal direction of the vehicle. The
controller area network (CAN) bus is used to aerodynamic drag is given by
communicate different components of the ACC
system with each other. Each component/node has 
CAN controller to pack and unpack the signals where is aerodynamic coefficient, is mass
to/from CAN messages; and CAN Transceiver to density of air, is frontal area of the car and is
transmit and receive CAN messages. Ultimately, the longitudinal vehicle velocity. The mass density of
designed ACC system is implemented in Simulink air may be taken as 1.225 (Rajesh, 2011).
using simscape blocks. The distributed adaptive The vehicle frontal area is calculated from its mass
cruise control sub systems are described in section according to equation (3) (Rajesh, 2011).
II; in section III the low level controllers in brake
control unit and engine control unit are developed.
Finally, implementation and result analysis are The longitudinal tire forces and is friction
described in section IV. forces from the ground that act on the tires. They are
calculated from each tires slip ratio, normal load on
the tires and friction coefficients of tire road
II. THE ADAPTIVE CRUISE CONTROL SUBSYSTEMS
interface. Assuming the vehicle won't do vertical
Distributed ACC system is consisted of a serious of motion, the net force in the vertical direction is
interconnected units, sensors and actuators. These
different components communicate through serial
communication network called Controller Area Considering load transfer effect
Network (CAN) (Di Natale et al, 2012). In this part
these units are designed. Different assumptions
considered in each component are described.
A. Vehicle Modeling Taking a moment about contact point of front tire
Vehicle has different parts such as diesel engine,
transmission unit, and wheels. The model of the
vehicle is complete representation of these parts. Substituting equations (3, 4, 5, and 6) in equation
Distributed Adaptive cruise control is an advanced (7) normal loads on tire will be:
longitudinal control of the vehicle which involves
longitudinal dynamics of the vehicle (Rajamani,
2011). Longitudinal dynamics of a vehicle has two
parts: power train dynamics and vehicle dynamics where
(Rajesh, 2011). The longitudinal power train of a  is vehicle length
vehicle is consisted of diesel engine, transmission  is distance from front wheel to center
and wheels along with engine speed sensors, wheel of mass along longitudinal direction
speed sensors, vehicle speed sensor and torque  is acceleration due to gravity
actuators. Vehicle dynamics quantifies the influence  is vertical distance from vehicle center
of longitudinal tire forces, aerodynamic drag forces of mass to floor
and gravitational forces on the vehicle. These  is normal loads on front tires
longitudinal forces are indicated in Fig.2. (Abe,  is normal loads on rear tires
2009).

Fig.2.Some the forces acting on a vehicle


Force balance along the longitudinal direction yields Fig.3.The front and rear tire behavior
(Abe, 2009)


54
Distributed Adaptive Cruise Control
ICAST-2017

The tire longitudinal slip ratio is computed from Wheel speed sensors
wheel dynamics and it is given by equation (10). Engine speed
Vehicle speed
Brake actuators
where is longitudinal slip, is effective radius. If
friction coefficient of the tire-road interface is
Car drive train unit
assumed to be 1 and the normal force is assumed to
be a constant, the typical variation of longitudinal
tire force as a function of the slip ratio is shown in
Fig 3. During braking the slip ratio might not be Wheel speed
Switch signal
small to adopt the linear small slip region for Engine speed
Brake torque
modeling tire force. So, the non-linear tire model Vehicle speed
called The Pacejka “Magic Formula" is used. It is
given as in equation (11, 12)
CAN Node 4

where coefficients are the parameters that


characterizes the tire and developed from the
experimental data. In paper, the tire is assumed to
have the characteristics shown in Fig. 3.
Fig.4.The car drive train unit:
The effect of power train components are included
to the model of the car by considering their effects Arrow direction indicates the direction of flow of signal.
through wheel dynamics; as the wheels are derived Written in the text box is the name of the signal
by the torque from engine (Levent et al, 2012). In The car drive train unit collects data from wheel
this case, the vehicle mechanical drive-line is speed sensor on each wheel, the engine speed sensor
modeled as a rigid body, with losses modeled by and vehicle speed sensor. The car drive train unit
efficiency. Diesel engine is modeled using a static has components that are responsible for driving and
torque map that defines the relationship between braking the car as requested by engine control unit
inputs of throttle position and engine speed and and brake control unit respectively. So, it receives
the output, engine torque . The throttle input throttle position signal from engine control unit and
comes from either the driver pressing gas pedal the brake torque needed to brake each wheel from
during manual operation or from the engine control brake control unit. It has torque actuators on wheel
unit during autonomous operation. The torque that process the received torque signal and brake the
output from the static map of the diesel engine is car. The throttle position signal controls the engine
transmitted to the wheels through the drive-line as torque directly according to the engine map while
torque controlling engine speed indirectly.
B. ACC Module
where represents the effect of mechanical derive The primary function of the ACC module is to
line. The dynamic equation of the wheel is process the radar information and determine if a
computed as follows by taking the moment about target vehicle is present or not. Unlike pulsed radar
center of the wheel systems that are commonly seen in the defense
industry, automotive radar systems often adopt
frequency modulated continuous wave (FMCW)
The vehicle body subsystem is termed as car drive radars. Compared to pulsed radars, FMCW radars
train unit in this work. The block diagram indicating are smaller, use less power, and are much cheaper to
the signals that the car drive train unit receives and manufacture. The radar occupies the band around
the data that it transmits on a CAN bus is shown 77GHz and it is mounted on the front of the ACC
Fig. 4. car (Camilla et al,2009).To measure the range and
Doppler, FMCW radar typically performs the
following operations (Atayants et al, 2014):
 The waveform generator generates the
FMCW signal.
 The transmitter and the antenna amplify
the signal and radiate the signal into space.
 The signal propagates to the target, gets
reflected by the target, and travels back to the
radar.

55
Distributed Adaptive Cruise Control
ICAST-2017

 The receiving antenna collects the signal. end


 The received signal is de-chirped and end
saved in a buffer Algorithm 1: ACC state algorithm
Once a certain number of sweeps fill the buffer, the
Fourier transform is performed in both range and Fig. 6 shows the block diagram of the ACC unit
Doppler to extract the beat frequency as well as the subsystem. This subsystem is interfaced with radar
Doppler shift. Buffering is needed, because the sensor that collects the speed and range of target
Doppler frequency is indistinguishable from the vehicle. This subsystem receives and transmits
beat frequency in single sweep. One can then different signals as indicated in the text box on the
estimate the range and speed of the target (Camilla fig. 6. It receives the ACC car speed from brake
et al, 2009) using equations (16, 17). control unit; the switch signal, the desired time gap
and desired speed signal from user interface unit.
Processing the received data, it computes the ACC
state of the car. ACC unit transmits the reference
speed, the target speed, target range and brake
signals on the CAN bus.

Radar

Fig.5. The FMCW modulation wave form (Camilla et al,


2009) ACC unit

Target speed Switch signal


where is relative velocity given by equation (16),
Ref speed ACC car speed
is carrier frequency, is speed of light, is Brake signal Desired time gap
modulation bandwidth, is range given by equation Target range Desired speed
(17) and is sweep time.
CAN Node 1

ACC determines the state in which ACC car has to


be using the following algorithm

Result: ACC states


if ACC switch is ON then Fig.6.The ACC unit subsystem:
Start the radar
if radar is ON then Arrow direction indicates the direction of flow of signal.
Read and R Written in the text box is the name of a signal
Read ACC car speed C. Brake Control Unit
while set do
if R≥200m then The brake control unit is the ACC subsystem
Set ACC car speed to responsible for braking the car without locking its
Brake signal OFF wheels. It continuously monitors the slip ratio of
End each wheel. The wheels rotate with initial angular
if R≤ 200m and ≤ then speeds that correspond to vehicle speed before
brakes are applied. The brake control unit monitors
Set ACC car speed to the wheel angular speed measured by wheel speed
Brake signal ON sensors and vehicle speed from speed sensors. Then
End it calculates the slip ratio using equation (18).
else
Set ACC car speed to
Brake signal OFF
End
End
end Slip equals one when the wheel is locked as wheel
else
is no longer rotating. A desirable slip value in this
Radar is not functioning
Exit
paper is 0.2, which is a peak value on the tire

56
Distributed Adaptive Cruise Control
ICAST-2017

characteristics. It is derived from fig. 3. At this slip,


adhesion between the tire and road is maximized
while the stopping distance is minimized due to
presence of friction. Brake control unit
To control the rate of change of brake pressure, the
model subtracts actual slip from the desired slip and
feeds this signal into the hydraulic low level Wheel speed signal Brake torque signal
controller. The output of hydraulic controller is an Vehicle speed signal
ON/OFF signal. The response of hydraulic system Brake signal
to ON/OFF signal looks like in fig. 7.

CAN Node 2

Fig.7. The front and rear tire behavior


Fig.8.The Brake control unit subsystem:
A system with such response to step input can be
modeled as a lag system of first order with transfer Arrow direction indicates the direction of flow of signal.
function as in equation (20). Written in the text box is the name of a signal
D. Engine Control Unit
The output torque and engine speed of the diesel
where is gain, is system lag. The output of engine of the vehicle depends on the throttle
hydraulic controller is a filtering rate. The model position . The engine control unit receives ACC
then integrates the filtered rate to yield the actual car speed from brake control unit and the needed
brake pressure. Multiplying piston area with the reference speed from ACC module to determine the
brake pressure gives the braking force. The brake acceleration demand of the vehicle. This unit is the
torque is computed as the multiplication of wheel heart of ACC system as it is responsible to provide
effective radius and braking force. the required demand in terms of throttle position to
The brake control unit: responsible for braking the the engine. PI control is used in order to make the
car without skidding, receives brake signal from ACC car speed to track the reference speed. One has
ACC unit, vehicle speed from vehicle speed sensor, to carefully tune the gains of this controller in order
and wheel speed signal from wheel speed sensors. to avoid the saturation problem as the engine
Vehicle speed and wheel speed signals are truncates any value of throttle position above one to
transmitted over CAN bus by car train subsystem. one and any value below zero to zero. In the
Fig. 8 indicates block diagram of brake control unit meanwhile, the ACC car speed has to track the
subsystem. Having received those data, the brake reference with very low deviation. The output of
control unit computes ACC car speed and brake this control is converted to equivalent engine speed
torque signal. using the look up table.
The other and very important part of the engine
control unit is the throttle controller part which is
responsible to bring the actual engine speed to the
desired engine speed computed from acceleration
demand. PI controller is used to make the engine
speed to converge to the desired engine speed. The
gains of the PI controller have to be tuned carefully
to avoid saturation problem. Being responsible for
computing the throttle signal that controls the
engine torque and speed, the engine control unit
requires the engine speed, ACC car speed and
reference speed data. It receives engine speed from
car drive unit, ACC car speed from brake control
unit and reference speed from ACC unit subsystem.
The engine control unit transmits throttle signal
over a CAN bus. Indicated in fig. 9 is the block
diagram of engine control unit subsystem.

57
Distributed Adaptive Cruise Control
ICAST-2017

Desired cruise speed button


Engine control unit Desired time gap button
Switch ACC button

User interface unit


Throttle position signal ACC car speed
Ref speed
Engine speed Desired time gap
Target speed
Desired cruise speed
Target range
Switch signal
CAN Node 3 ACC car speed

CAN Node 5

Fig.9.The Engine control unit subsystem: Arrow


direction indicates the direction of flow of signal. Written
in the text box is the name of a signal Fig.10.The User interface unit subsystem: Arrow
direction indicates the direction of flow of signal. Written
E. User Interface Unit
in the text box is the name of a signal
The primary function of the user interface unit is
F. Communication Protocol
to manipulate the Cruise Switches and send their
information to the ACC unit and car drive train The Controller Area Network (CAN) is an
unit. The cruise switches are the ACC switch, time automotive standard network that utilizes a 2
gap button and desired speed button. The ACC wire bus to transmit and receive data. Each node on
switch is used to turn on or off the ACC i.e. it is the network has the capability to transmit 0 to 8
used to switch between manual driving and adaptive bytes of data in a message frame. A message frame
cruise control mode of driving. The time gap button consists of a message header, followed by 0 to
is used to set the desired time gap between ACC car 8 data bytes, and then a checksum. The message
and Target vehicle. The desired speed button is used header is a unique identifier that determines the
to set the desired cruising speed of the car. The user message priority (Di Natale et al, 2012). Any node
interface unit also displays the target vehicle speed, on the network can transmit data if the bus is
the range from target vehicle and ACC car speed free . If multiple nodes attempt to transmit at
with different text messages for the driver so that the same time, an arbitration scheme is used to
the driver has information regarding the state of the determine which node will control the bus. The
ACC system. Fig. 10 shows the block diagram of message with the highest priority, as defined in
the User interface unit. This unit sends the switch its header, will win the arbitration and its
signal, the desired time gap and desired speed signal message will be transmitted. The losing message
over a CAN bus. It receives ACC car speed from will retry to send its message as soon as it detects a
brake control unit, Target car speed and Target bus free state (Di Natale et al, 2012).ACC system
range from ACC unit subsystem. has five CAN nodes that communicate over CAN
bus. CAN node 1 for ACC unit subsystem, CAN
node 2 for brake control unit subsystem, CAN node
3 for engine control unit subsystem, CAN node 4
for the car drive train subsystem and CAN node 5
for user interface unit subsystem. On the car there
are four wheel speed sensors, four wheel actuators,
and one car speed sensor, one engine speed sensor
and one radar sensor. The ACC unit subsystem
receives actual car speed from brake control unit
and the relative speed of target vehicle and relative
range to target vehicle from the radar and computes
the desired reference speed and transmits brake
signal when needed. The reference speed is received
by the engine control unit and brake signal is
received by the brake control unit. In addition,

58
Distributed Adaptive Cruise Control
ICAST-2017

No Message Name Node Number ID Length


1 Reference speed 1 0 76 bits 5 ms 2µs 0.19ms
2 Brake signal 1 100 52 bits 5 ms 2µs 0.12ms
3 Switch signal, set 5 21 92 bits 5 ms 2µs 0.27ms
4 ACC car speed 2 600 76 bits 5 ms 2µs 0.19ms
5 Vehicle speed 4 700 76 bits 5 ms 2µs 0.19ms
6 Throttle position 3 800 76 bits 5 ms 2µs 0.19ms
7 Engine speed 4 900 76 bits 5 ms 2µs 0.19ms
8 Target range 1 1000 76 bits 5 ms 2µs 0.19ms
9 Target speed 1 1100 76 bits 5 ms 2µs 0.19ms
10 Brake torque front left 2 1200 76 bits 5 ms 2µs 0.19ms
11 Brake torque front right 2 1300 76 bits 5 ms 2µs 0.19ms
12 Brake torque rear left 2 1400 76 bits 5 ms 2µs 0.19ms
13 Brake torque rear right 2 1500 76 bits 5 ms 2µs 0.19ms
14 Wheel speed front left 4 1600 76 bits 5 ms 2µs 0.19ms
15 Wheel speed front right 4 1700 76 bits 5 ms 2µs 0.19ms
16 Wheel speed rear left 4 1800 76 bits 5 ms 2µs 0.19ms
17 Wheel speed rear right 4 1900 76 bits 5 ms 2µs 0.19ms
Table I. Static Properties of Can Messages In Acc System

engine control unit receives engine speed and car the message. Messages with smaller ID will have
speed from the car drive train and brake control unit higher priority. Each message have static properties
respectively. Throttle position signal, and brake which comprise message priority , message
torque signal for each wheel of the vehicle is period , message maximum transmission time ,
received by the car drive train unit subsystem. The message arrival time , (Robert et al, 2007). The
brake control unit computes car speed and wheel maximum transmission time , is determined by the
brake torque signal while it receives wheel speed number of data bytes, bit stuffs and inter frame
and car speed from car drive train. The user space, and it is computed by equation (21).
interface unit receives signals from ACC unit and
brake control unit and displays the information
about the traffic and information about state of the where is number of byte in data field of message,
ACC car for the driver. Fig. 11 is the block diagram is the time it takes to transmit one bit. The static
of the system. Written in the text box are the signal properties of CAN messages used in this work are
names with the arrows direction indicating the described as in table I.
direction of signal flow. In each CAN node there are
Apart from CAN messages with message ID of 100
two sub parts: the CAN controller and the CAN
and 21, all other CAN messages have transmission
transceiver. CAN controller is responsible for
time of 0.19 ms. So, for messages with message ID
packing and unpacking of the CAN messages. CAN
greater than 100, the maximum blocking time by the
transceiver is responsible for transmitting and
lower priority message occupying the bus is 0.27
receiving messages on the CAN bus. CAN receiver
ms. The blocking time by the lower priority
doesn't block messages rather receives all the
message is given by:
messages which are latter ignored by CAN
controller if it is not the needed one. CAN bus used
in this work is the high speed CAN bus with 0.5
Mbps bus speed. Each CAN message transmitted = set of messages with lower priority than
over a CAN bus is required to have unique ID. The message if a message is the lower priority
standard ID which is 11 bits in length is adopted for message, is equal to zero).
each message. When a CAN transmitter attempts to The ACC system is continuous system while the
transmit messages, it has to first detect if the bus is controller area network is a discrete system.
idle or not. If two/ more messages from the same Discrete system manipulates discrete signal; so, the
node or two or more nodes start to transmit continuous signals from the plant have to be
messages simultaneously, the bus arbitration based sampled and feed to CAN nodes. A zero holder
on message ID is triggered. The message with high sampler with sampling time of 5 ms is used to
priority which is a message with small ID will be sample the continuous signal from the ACC sub
transmitted winning the arbitration process. CAN systems.
messages are queued in priority queue in the nodes All the CAN messages packed by CAN controller
before being transmitted. The priority in the queue are sent over a CAN bus by a CAN transmitter
is based the value of identifier field in the header of periodically. The period of each message is assumed

59
Distributed Adaptive Cruise Control
ICAST-2017

to be equal and it is 5 ms. The CAN bus adopted for gap. It shouldn't be affected by the transition effects
this system has a bus speed of 500 Kbps. With this from one mode of operation to another.
rate of bus speed, the time that it takes to transmit The first part of the throttle control unit is the
one bit will be 2 µs. The message length is acceleration demand part. The acceleration demand
calculated by adding the number of bits in the is computed from the reference speed and the ACC
header of the message to the number of data bits car speed. This demand quantifies the amount of
from the data field of the message. The message engine speed needed for the car to track the
transmitted by CAN transmitter is received by CAN reference speed and it is computed by PI control. It
is possible to call this a vehicle speed controller in
which the output is the acceleration demand.

Brake
Engine control unit
User
control
Interface
Unit unit

CAN CAN Node CAN Node


Node 5 3 2

Fig.12. Acceleration demand


CAN BUS
The second part of throttle control is the throttle
position controller which controls engine speed and
engine torque. The reference speed required of the
CAN CAN Node
engine comes from the first part of throttle control
Node 4 1 unit. This reference speed is computed from the
acceleration demand using a look up table that
relates the demand acceleration to the engine speed
map. Again PI controller is used to make the engine
Car drive ACC Unit
train unit
track the desired engine speed. The output of the
throttle controller is the throttle position. Throttle
position controls engine speed directly and engine
Fig.11. Block diagram of the Adaptive cruise torque indirectly.
control system: Arrow direction indicates the
direction of flow of signal.

receiver in each CAN nodes. CAN receiver samples


the received messages with sampling time of 5 ms.

III. INPUT SATURATION AND CONTROL DESIGN


Fig.13. Throttle controller
In order for the designed ACC system to meet its
requirements, low level controllers are needed in B. Limited Brake Control
engine control unit and brake control unit. The
Brake control is part of brake control unit. It is
control design has to take into account saturation
responsible to brake the car without locking the
problem and tracking requirements. The ACC car
wheel lock. Its primary purpose is to avoid the
speed has to track the desired speed with very low
wheel lock while braking. This controller is a
steady state error and small over shoot. Normal
simple on/off control depending on the value of
ACC systems are not functional for low speed
relative slip. The brake torque is computed from
scenarios i.e. below 30Mph. In this work low speed
brake pressure and wheel radius as:
scenarios are also given a consideration.
A. Throttle Control
Throttle control is the low level controller in the where is brake pressure and is brake pressure
engine control unit. It is required of to make the car rate. The brake force is then found as:
speed to follow the desired reference speed. It
should work in any scenario of ACC states: speed
control scenario where the car cruises at the set where A is cylinder area.
speed by the driver, time gap control where the car
has to follow the target vehicle with the given time

60
Distributed Adaptive Cruise Control
ICAST-2017

where is brake torque and is effective wheel low level hydraulic brake controller, the wheel
radius. dynamics and longitudinal vehicle dynamics. CAN
controller and CAN transceiver are implemented by
the vehicle network block while the car drive train is
IV. IMPLEMENTATION AND RESULT ANALYSIS
implemented by the simscape block. The sensors
The designed Distributed ACC system is and actuators are implemented by blocks from
implemented and analyzed using simulink block, simscape. The low level controllers in the engine
vehicle network block and Simscape block for control unit are implemented by simple simulink
different scenarios. User defined matlab functions blocks. Different figures are plotted in order to
are used to implement the ACC state algorithm, the analyze the performance of the designed system.

Fig.14. Speed control scenario at 40 Km/hr with target 100m away and desired time gap set to 0.9 seconds.

Fig.15. Time gap control scenario with constant relative range of 20 m between cars and both cars moving at the same
speed

61
Distributed Adaptive Cruise Control
ICAST-2017

Fig.16. Time gap control scenario with variable relative range between cars

Fig.17. The mixed scenario: speed control at 40 Km/hr with target car 44.4m away and time gap control of 0.9 sec with
target car 20m away. In both cases the relative speed between cars is assumed to be zero.

large distance between ACC car and target


There are so many kind of interaction scenario of vehicle, ACC car is controlled by setting desired
the ACC car with traffic condition on the road in a speed i.e. the ACC car is in speed control mode.
similar lane. When the target car is out of Fig. 14 shows the speed control scenario at 40
detection range of radar sensor or when there is while the relative distance
between the cars is 100 m with the desired time and speeding down making the relative speed
gap set to 0.9 seconds. between cars varying. Fig. 16 illustrates the
scenario in which the relative speed between cars
The time gap control mode is used when the target varies by 0.1 Km/hr and the target car is initially
range is smaller. In this case the car tries to 20m away from the ACC car.
maintain the desired time gap rather than the set
The other scenario is the scenario which involves
desired speed. Fig. 15 describes the time gap
both speed control and time gap control. Fig. 17
control scenario with desired time gap of 0.9 sec
describes the speed control scenario and time gap
and assuming the cars are moving with similar
control scenario in which the ACC car is
speed and the target car is 20m away from ACC
controlled to keep the desired time gap of 0.9
car. There are scenarios in which the relative
seconds to a target vehicle which is 20m away for
range can be varying i.e. the front car speeding up

62
Distributed Adaptive Cruise Control
ICAST-2017
some periods and then enters speed control mode where U is network utilization rate, N is number
when the target range grows to 44.4m. of messages, is message transmission time and
is message period. When message period is
When CAN networks are used in the
increased, the network utilization rate is reduced
implementation of the closed loop system; as a
and this has a negative effect on the performance
result of the limited network bandwidth, two
of the designed control system as shown in fig.
sources of delays may occur. These are sensor to
19. In the figure the speed of the ACC car when
controller delay and controller to actuator delay.
the message period is 5 ms is compared to the
Any controller computational delay can be
speed when the message period is 100 ms.
absorbed into either delay without loss of
generality. The delays can affect the transient The network utilization rate when the message
response behaviors in a control system. The period is 5 ms is 65% while the network
general system performance as described by the utilization rate when the message period is 100 ms
criteria for step response may be degraded. This is 3.25%. The response of the system became
effect may include the increase of both overshoot oscillatory with the increased sampling time.
and settling time of the system step response. The Similarly, very high network utilization rate
delays can also affect the stability of a system and degrades the performance of the control system
cause the system to become unstable. Illustrated due to over loading of the network. The number
in fig. 18 is the comparison of speed of ACC car of messages to be transmitted over CAN network
with and without controller area network. As it is plays crucial role in choosing suitable message
clear from the figure, it is the transient behavior of period. The other important parameter in network
the system that is highly affected. Speed control based control system is the delay of messages.
mode is less affected than the time gap control The maximum delay a message my experience is
mode as the transient behavior is only for shorter calculated from the CAN Log data. From CAN
period. log data it is only possible to determine the
maximum delay, mean delay of the messages are
The CAN transmitter used in this work samples
difficult to determine. The maximum root mean
messages with a sampling period of 5ms.
square delay for a CAN network with sampling
Sampling period controls the network utilization
messages at a period of 5 ms and CAN bus speed
rate computed by:
of 500Kbps is 0.0048 seconds.

Fig.18. Speed comparison: The speed of ACC car when CAN bus is in the control loop is compared to Speed of ACC
car when components are connected by ideal connection

63
Distributed Adaptive Cruise Control
ICAST-2017

Fig.19. The response of control system when sampling messages at different period. The speed at message sampling
period of 5ms and the speed at message sampling period of 100 ms are plotted

V. CONCLUSION AND FUTURE WORKS REFERENCES


In this project the adaptive cruise control is 1. Levent Guvenc, Et Al. (2012), Cooperative
designed and implemented for different scenarios Adaptive Cruise Control Implementation Of Team
including low speed cases. The vehicle under Mekar At The Grand Cooperative Driving
study is modeled using drive train dynamics and Challenge, IEEE Transactions On Intelligent
vehicle dynamics. Low level controllers are Transportation Systems, IEEE.
designed to control throttle position and brake 2. Idar Petersen, Et Al. (2001), Wheel Slip Control In
pressures respectively. In the future a complete ABS Brakes Using Gain Scheduled Constrained
model of the vehicle can be used to design ACC LQR, European Control Conference.
system for a stop and go scenario. To have a 3. Arvind F., Et Al. (2011), Cruise Control Operation
from Zero to Preset Speed-Simulation and
complete model of car one has to include the
Implementation, International Journal of
rolling resistance and road level effects in
Information and Education Technology.
longitudinal vehicle dynamics. In addition to this 4. Rajesh Rajamani (2011), Vehicle Dynamics and
the transient behavior of the tire should be Control, Springer
included in wheel dynamics. 5. Camilla Karnfelt, Et Al. (2009), 77 GHz ACC
Radar Simulation Platform, ITST 2009.
6. Masato Abe. (2009), Vehicle Handling Dynamics
1st Edition Theory and Application, Butterworth-
Heinemann 2009
7. Marco Di Natale, Haibo Zeng, Paolo Giusto.
(2012), Understanding And Using The Controller
Area Network Communication Protocol: Theory
And Practice, Springer.
8. Boris A. Atayants, Viacheslav M. Davydochkin,
Victor V. Ezerskiy, Valery S. Parshin, Sergey M.
Smolskiy. (2014), Precision FMCW Short-Range
Radar for Industrial Applications, Artech House
9. Robert I Davis, Et Al. (2007), Controller Area
Network (CAN) Schedulability Analysis, Real-
Time Systems, Springer

64
Design of Controller via Feedback Linearization for a Non-linear Conical Tank ICAST-217

Design of Controller via Feedback Linearization for a


Non-linear Conical Tank
Dr.K.V.L.Narayana1, Kena Likassa Nefabas 1
1
Department Of Electrical Power And Control, Adama Science And Technology University, Adama City, Postal
Code:1888, Ethiopia,
Email: 1KVLNARAYANA@YAHOO.CO.IN , 2KENA0912@GMAIL.COM,

Abstract — Nonlinear process control is a difficult task in process industries. They exhibit nonlinear
behavior, time variance and delays between inputs and outputs. This paper aims at implementing
feedback linearization technique in which a nonlinear system can be converted into a linear system.
Feedback linearization technique is completely different from conventional linearization which is
carried out around nominal operating point. Feedback linearization transforms the entire nonlinear
system to a linear system. In this paper, feedback linearization is used to linearize a nonlinear tank
and then state feedback control is implemented and compared with IMC-PID controller designed via
linearization around operating points.
Key words- Conical tank, feedback linearization, state feedback control, IMC, IMC-PID
From which it is clearly evident that the system is
I. INTRODUCTION nonlinear. After substitution the parameters of
conical tank the equation (1) is simplified as
Many process industries use conical tanks because
of its shape contribute to better drainage (Gunaselvi,
2013). Conical tank is highly nonlinear due to
varying cross sectional area and dependence of out To control a nonlinear process it is common practice
flow rate on square root of liquid height (h) in the to linearize the system around operating points and
tank. Conical tank is widely used in paper schedule the controller gains according to the
industries, petro chemical industries and water operating points. The conventional linearization for
treatment plants etc (Madhubala, 2004; Christy & the system around the operating point at a height ‘h’
Dinesh Kumar, 2014). The control of liquid level in results into a transfer function as shown in equation
tanks and flow of the tank is a basic problem in 3.
process industries. A level that is too high may
upset reaction equilibrium, cause damage to
equipment, or result in spillage of valuable or
hazardous material. If the level is too low, it may
have bad consequences for the sequential operations
(Madhubala, 2004; Christy & Dinesh Kumar, 2014).

II. PROCESS DESCRIPTION & MATHEMATICAL


MODELING
The schematic diagram shown in Fig. 1
(Madhubala, 2004; Christy & Dinesh Kumar, 2014)
shows various parameters of a conical tank.
Where
Fin=inflow rate (maximum of 400 cm3/sec) Fig. 1. Schematic diagram of the process
H=Maximum height (73cm)
R=Maximum radius of conical tank (19.25cm)
Kv =Valve constant=55 cm2/sec
A=Cross sectional area
According to the mass balance equation the tank Where
model is
(4)
(1)
and

65
Design of Controller via Feedback Linearization for a Non-linear Conical Tank ICAST-217

Where
III. THEORY OF IMC-PID CONTROLLER
In IMC-PID controller design the process model is
explicitly used in the control system design
procedure. The standard feedback structure uses the Where ‘’ is the tuning parameter of IMC which
process model in an implicit fashion, that is, PID varies the speed of response and ‘n’ is order of
tuning parameters are “tweaked” on a transfer process.
function model, but it is not always clear how the
process model effects the tuning decision B. Design of IMC-PID controller for delay free
(Kongratana,2012; Xiang et al., 2013). In the IMC first order process
formulation, the controller, q(s), is based directly on Consider a first order process
the “good” part of the process transfer function
(Kongratana,2012; Xiang et al., 2013). (9)

The IMC formulation generally results in only one (10)


tuning parameter, the closed loop time constant (,
and the IMC filter factor). The PID tuning
And from equation (7)
parameters are then a function of this closed-loop
time constant. The selection of the closed-loop time (11)
constant is directly related to the robustness
(sensitivity to model error) of the closed-loop
system. Arranging the above equation as

A. The equivalent feedback form to IMC (12)


IMC can be arranged in the form of standard
feedback form as shown below in Fig. 2. And comparing it with standard PI controller
.
d(s)

+
+ And
r(s) I. + h
_
C. Feedback linearization of conical tank
Feedback linearization is an approach to nonlinear
control design that has attracted lots of research in
Fig. 2. Feedback Equivalent structure of IMC recent years (Lichun et al., 2006).The central idea is
to algebraically transform nonlinear systems
Where dynamics into fully linear ones, so that linear
d(s) = disturbance control techniques can be applied. This differs
gp(s) = actual process entirely from conventional (Jacobean) linearization,
gp(s) = estimated process because linearization is achieved by means of
q(s) = IMC = gp(s) feedback, rather than by linear approximations of
gc(s) = feedback controller the dynamics. The implementation of feedback
linearization for a nonlinear conical tank is
So the standard feedback controller equivalent to explained as follows.
IMC structure is
Suppose if the input flow rate (Fin) is selected as

In practical to minimize the model mismatch a low Where ‘v’ is a synthetic input.
pass filter is added in series with the internal model
controller. So the controller transfer function is From equation (1)
(16)
66
Design of Controller via Feedback Linearization for a Non-linear Conical Tank ICAST-217

and i.e The final value of h(t) is .

This ensures that desired set point is achieved. The


selection of constant K affects the speed of response
Now the system is linear w.r.t input v and output h. of process. But the actual system cannot take input
So a linear controller can be designed which gives as .So using equation (15) actual required
the value of synthetic input v. But the actual system physical input is derived from the synthetic input
‘ ’
takes the physical input as but not in terms of
synthetic input v. So using the equation (15) (23)
will be calculated from the synthetic input and this
will be fed to the actual system. The overall After substituting the parameters of conical tank
working principle is shown in Fig. 3.

E. Design of IMC-PID controller


CONVERTIO
N OF CONICAL From section (3) the transfer function of conical tank
SYNTHETIC TANK around operating point h is
I/P (v) INTO
PHYSICAL
I/P( ) h

Where
LINEAR
CONTROLL
ER and
v
Substituting the parameters of conical tank of
interest
Fig. 3. Schematic of control strategy via feedback
linearization (25)
D. Design of full state feedback controller and
In this section the design of Full State Feedback
controller (FSF) is illustrated. From the equation (26)
(16) the system in hand is now linear between
And from equations (13) & (14) the values of and
output and synthetic input . proportional and integral gains are
If the synthetic input is selected as
(18)
&
Where
Substituting and values from equations (23)
And & (24) will give

K=Positive constant
and (28)
Now Equation (16) became
(19) IV. RESULTS AND DISCUSSION
Taking Laplace transform on both sides In this section performance of IMC-PID and Full
State Feedback (FSF) controllers is compared. The
(20) performance is compared in terms of set point
tracking and disturbance rejection. Fig. 4 shows
controller performance for a set point of 40 cm.
(21)

Using final value theorem


(22)
67
Design of Controller via Feedback Linearization for a Non-linear Conical Tank ICAST-217

45
controller designed via linearization around
40 operating points.
35

30
60
(levelm)

25

(c m)
20
40
IMC-PID
SETPOINT

Lev el
15 FSF

10 20 IMC - PID
Set Point
5

0 0
0 50 100 150 200 250 300 350 400 450 500 0 100 200 300 400 500 600 700 800 900 1000
Time(s)
Time (s)
Fig. 4 Comparison of IMC-PID vs FSF for a Set Point of
40 cm
60
Fig. 5 shows performance of controllers against set
point changes

(c m)
40

Lev el
Table I. Comparison of Time domain characteristics Set Point
20
FSF
Type of Rise Settling time Peak over
controller time(s) (s) shoot (%)
0
IMC-PID (=10) 83.64 217.18 0 0 100 200 300 400 500 600 700 800 900 1000
Time (s)
FSF (K=10) 32.34 48.05 0
Fig. 6. IMC-PID vs FSF against disturbance (A
Disturbance of 10cm and -10cm is introduced at 300s and
500s respectively)
50

45 REFERENCES
40 1. Gunaselvi Manoharl, Elakkiya V, Pearley Stanlel
35 and Sudha R. (2013). Neural Network Based Level
30 Control in Two Tank Conical Interacting System,
Level(cm)

25
Proceedings of International Conference on
Intelligent Systems and Control (ISCO 2013),
20
IMC-PID (pp.194-196)
15 SET POINT
FSF 2. Madhubala, T K., M Boopathy, Sara Chandra Babu
10 and T K.Radhakrishnan (2004). Development and
5 tuning of fuzzy controller for a conical level system,
0
Proceedings of international Conference intelligent
0 200 400 600 800 1000 1200 1400 1600
Time(s)
sensing and information Processing, (pp. 450-455).
Tiruchirappalli, India.
Fig. 5. Performance comparison against set point changes 3. Y. Christy and D. Dinesh Kumar (2014). Modeling
Figure 6 shows controller performance against and Design of Controllers for Interacting Two Tank
Hybrid Systems (ITTHS), International Journal of
disturbance. A disturbance of 10 cm and -10cm is Engineering and Innovative Technology (IJEIT), 3
introduced at 300s and 500s respectively. (7), 231-239
4. Kongratana V, Tipsuwanporn V and Numsoran A
VI. CONCLUSION (2012). IMC based PID controller design for
torsional vibration system, 12th international
Feedback linearization is a powerful tool which conference on control, automation and systems Jeju
eliminates the necessity of partial linearization Island,korea.
around operating points and thus scheduling of
5. Leijun Xiang, Yinyin Dai and Xiaofen Lu (2013).
controller gains. To increase the robustness of the
Novel IMC-PID Controller Design and Parameter
controller integral of error can be added to the
Tuning with Improving Control Performance for the
synthetic input (v). Feedback linearization of SOPTD Processes 10th IEEE International
conical tank gave excellent servo and Regulation Conference on Control and Automation (ICCA)
control performance when compared to Hangzhou, China.
conventional linearization. The response is smooth
as well as fast when compared to convention
68
Design of Controller via Feedback Linearization for a Non-linear Conical Tank ICAST-217

6. Lichun Li, Zhufeng Wang, Jingping Jiang, Stabilization Scheme for Three-Phase Grid-
Xiaoming Yu (2006). Robust Design for Network- Connected Photovoltaic Systems, IEEE Journal of
based Control System with Random Time Delay, Photovoltaics, 4(1), 423-431.
TENCON 2006. Proceedings of 2006 IEEE Region
10 Conference.
7. Mahmud M. A., Pota H. R., Hossain M. J., and Roy
N. K. (2014). Robust Partial Feedback Linearizing

69
Design of Controller via Feedback Linearization for a Non-linear Conical Tank ICAST-217

70
Image processing for Pick and Place Gantry Robot
ICAST-217

Image processing for Pick and Place Gantry Robot


1
yonatan Tasew, 2daniel Nigusse
1
Department of Electromechanical engineering, Addis Ababa Science and Technology University, Addis Ababa, Postal
code: 1230, Ethiopia.
yonatantasew@gmail.com, daniel925nigusse@yahoo.com

Abstract — In developed countries almost all manufacturing sectors employ automation and
computer aided manufacturing (CAM) to produce reliable and accurate products that can compete in
the market. But our country’s manufacturing sectors seems to struggle with that concept and still
chooses to relay on the good old man power that is unpredictable, inaccurate, slow, and expensive to
work with. So in this paper an attempt is made to apply some image processing techniques to improve
hoisting mechanisms of an industry for a better handling and transporting of products and raw
materials by using gantry robot.
Key words- CAM, Image Processing, Gantry Robot
automation reduces the total production cycle
I. INTRODUCTION time, and second, it helps smooth flow of raw
In the manufacturing industries a lot of efforts are materials and finished products to their desired
expended in material handling tasks. This location with little or no human interference
situation has produced a significant increase in (Fredonia Focus).
the demand of material handling products that Gantry robots are electromechanical devices that
allows reducing these efforts, decreasing the can perform different tasks that are difficult,
consumed time and the costs of this activity. dangerous, repetitive or dull for human beings
Materials handling products and systems are (Abdela et al., 2016). They are devices that can be
found in almost every manufacturing and programmed to follow a set of instructions to
distribution company and for an endless number perform certain tasks (Abdela et al., 2016). The
of goods. robot consists of a manipulator mounted onto an
Automation has brought several drastic changes overhead system that allows it to move across
in manufacturing over the past century. These horizontal plane (Abdela et al., 2016). Gantry
changes include increase in overall productivity robots are also called Cartesian or linear robots
and profitability of a manufacturing system. (Abdela et al., 2016). They are usually large
Development of electronics (transistors and systems that perform pick and place applications,
microchips) led to a jump in control technology but they can also be used in welding and other
and precision of various instruments (Butala et applications. Gantry robot systems provide the
al., 2002). This laid a path for efficient and cost advantage of large work areas and better
effective manufacturing processes (Butala et al., positioning accuracy (John Craig, 2005). Position
2002). Automation of manufacturing systems accuracy is the ability of the robot to place a part
requires integration of various fields such as correctly (Abdela et al., 2016). Gantry robots are
mechanical, control and electronic systems, and easier to program, with respect to motion, because
computers (Butala et al., 2002). Automation in they work with an X, Y and Z coordinate system
various forms constitutes the backbone of most (John Craig, 2005). Another advantage is that
major industries. It has become a significant they are less limited by floor space constraints.
component of defense, medical, aerospace and Image processing is a form of signal processing
automotive industries, material processing and for which the input is an image, a series of images
handling, manufacturing and consumer products or a video and the output could either be an image
to meet the increasing demand for the production or a set of characteristic or parameters related to
volume and product variety. In the manufacturing the image (Adrian Rosebrock, 2014). Most image
industry, automation in material handling has processing techniques involve treating the image
increased the overall profitability of the product as a two dimensional signal and applying standard
with an improvement in the quality and signal-processing methods using mathematical
productivity of the system (Fredonia Focus). This operations (Adrian Rosebrock, 2014).
is primarily due to two reasons. First,

71
Image processing for Pick and Place Gantry Robot
ICAST-217

II. DESIGN OF MECHANICAL AND COMPUTER from the sensor matches the desired data given by
SYSTEM the user. That could be simply switching one device
The design of mechanical system of the device is on or turn the other off and the like.
responsible for the support and housing structure. It We can divide this system in to three main
is where all the electric components, gripping tool categories
and actuators are secured and the translation of The graphical user interface (GUI):- is the part
motion takes place. of the computer program (code) that is directly in
The mechanical part of this device is very important contact with the user and gather information in the
and needs a lot of details because any failure in in form of numerical value.
this part will result in total loss of control and
stability. So it must be carefully designed and
installed without errors.
The translation motion is produced by attaching the
threaded bars to the motors and a fixed nut can be
dragged parallel to the axis of rotation with the
bearings on the iron bar effectively producing
sliding motion along the angle irons.

Fig 3. Example of a GUI


The embedded firmware: - is a compiled code
that is embedded on the microcontroller which is
written to receive the numerical data given by the
user with the help of the GUI and the sensor data so
as to compute the correct estimate of actuation
period. This code calculates the rotor position which
Fig. 1. Bearing mechanism is gathered from the rotor encoders and gives the
command to turn on the actuators until the desired
III. ELECTRICAL CIRCUIT position is mate.
A. Electrical Circuit Image processing- this part of the project
focusses on filtering out the desired objects or
Is responsible for the connection of actuators, signal
colors from a video sources (camera) and use the
conditioning and sensing devices, switching de-
coordinates gathered for gantry feedback.
vices, power supply, microcontrollers and such to
receive the commands from the graphical user inter-
face or the user and process it through the
microcontroller and by using relays and transistors
control amount of time needed to turn the motors on
or off and direction of rotation.

Fig 4. Shape detection demo (circle)


C. Control system
Fig 2. Driver ICs and Microcontroller
The Control system helps in regulating the
B. Computer Systems performance of the device more precisely and uses
A The computer system is used for creating algorithms to cancel out unwanted oscillations,
algorithms that can receive numerical data from the transient problems and puts the system in steady
user and the sensors to calculate and estimate what state position. For the porpoise of this paper a
standard PID controller algorithms are used.
the microcontroller should do next until the reading

72
Image processing for Pick and Place Gantry Robot
ICAST-217

2. Stanley Cobotics, “Stanley assembly”. Received


from http://www.stanleyassembly.com/home.aspx
3. Schmidt-Handling, “KOBOT”. Received from
http://www.schmidthandling.de/Kobot.htm
4. Shape detection using opencv and python. Recieved
IV. CONCLUSION from
http://www.pyimagesearch.com/2016/02/08/opencv-
In general the expected outcome of this machine shape-detection/
can be enormous. In the material handling
science, one of the proper way of handling 5. Siciliano_B.,_Sciavicco_L.,_Villani_L._ Oriolo
material is done by this robot. And also it+ can be _G.(2008).Robotics: modelling, planning and
used for precise manufacturing of products as this control. Springer,
device is computer controlled machine it allows
for a very precise machining capabilities that can 6. Craig, John.(2005). Introduction to robotics:
be used to manufacturing products. A great leap mechanics and control. Upper Saddle River, NJ,
towards the dominating technology it is USA:: Pearson/Prentice Hall,
undeniable that our workshops and most of our
industries are manually operated but the rest of 7. Abdela Kemal, Abdulhamid Talib, Abiy Amare,
the world is using computer controlled and Abraham G/medhin, Fitsum Solomon.( 2016):
automated systems. And this device can be the Design and modeling of Gantry robot (Unpublished
starting stone to launch a generation of workers bachelor’s thesis). Addis Ababa Science and
that narrow this huge technological gaps. It also Technology University
initiates workers to adopt and implement the
latest technologies, since the technology gap 8. Butala P., Kleine J., Wingen S., Gergs H.(2002,
between our country and the rest of the world is May). “Assesment of Assembly Processes in
European Industry”, Proc. 35th CIRP International
large it will take some time to get in the front part
Seminar on Manufacturing Systems, Seoul, Korea.
of the creating and inventing of new technology,
but until such time comes we can adopt the 9. Yim M., Wie-Min Shen, Salemi B., Rus D., Moll
technologies to our specific requirements and M., Lipson H., Klavins E., Chirkijan, G. S.(2007).
solutions meaning cheaper and simpler designs of “Modular Self Reconfigurable Robot systems –
the adopted model. Grand Challenges of Robotics”, Robotics and
Automation Magazine, Vol. 14, Issue 1, pp. 43-52,.
REFERENCES
1. Fredonia Focus, “Material handling market research
Report”, Word material handling products. Received
from http://www.marketresearch.com/

73