Sei sulla pagina 1di 39

Almadar Journal for Communications, Information

Technology, and Applications


Vol. 03, No. 01, April 2016

AJCITA R EVIEWERS OF THIS I SSUE


Almadar Journal for Communications, Information Tech- Ashraf Saleem, Hassan Yousef, Azeldien Kinsheel, Ibrahim
nology, and Applications (AJCITA) (ISSN 2313−156X), Almerhag, Ahmed Alkilany, Khawla Alnajjar, Elhusain Saad,
(Local registration number at the Libyan National Li- Eshag Larbah, Idris El-Feghi, Mustafa Abuzaraida, Zia Nadir,
brary 369/2014) is published semi-annually by the Research Mohammed Bait-Suwailam.
and Development Office, Almadar Aljadid Co. Address:
Gurji−Tripoli−Libya, P.O.Box 83792, +218 91 919 0500, Ext. C OVER PAGE D ESIGN
3520, ajcita@almadar.ly, www.almadar.ly, www.almadar-rd.ly. Mohamed E. Glaiow
AJCITA aims to provide high-level scientific papers with Ahmed A. Aljarray
value in academic research and industry. The published papers
can be:
S UBMISSIONS
• Original research articles with clear contributions,
• Review articles,
The journal welcomes submissions of original research work
• Application articles with industrial values,
in the area of telecommunications, information technology
• Invited articles.
and applications. A 500 LYD is granted for accepted papers.
Best paper award every year will be granted 5000 LYD. All
communications between editorial board and correspondent
D IRECTOR OF J OURNAL
authors will be through ajcita@almadar.ly. The submitted
Abdulla A. Abouda, Manager of Research and De- papers will be checked first by the editors to confirm that
velopment Office, Almadar Aljadid Co. Libya, E-mail: it follows the main journal requirements in terms of topics
a.abouda@almadar.ly. and style. If the paper passed the first check it will be sent to
at least three independent peer reviewers to give acceptance
E DITOR IN C HIEF statement about the paper. The reviewers’ identities will be
Mohammed S. Elmusrati, Professor and Head of Communi- masked from the authors.
cations and Systems Engineering Group, University of Vaasa,
Finland. He is also Professor at Electrical and Electronic En- A DVERTISING
gineering Department, University of Benghazi, Libya, E-mail: Advertising in AJCITA is accepted at the discretion of the
mohammed.elmusrati@uwasa.fi. publisher. In the next couple of issues accepted advertisement
will be published free of charge. All communications related
E DITORIAL B OARD to advertising should be directed to ajcita@almadar.ly.
Naser G. El−Tarhuni, Assistant Professor at Electrical
and Computer Engineering Department, University of Sultan C OPYRIGHT
Qaboos, Sultanate of Oman, E-mail: tarhuni@squ.edu.om. All rights reserved for AJCITA, however, abstracting is per-
Ali A. Ganoun, Assistant Professor at Electrical and mitted with credit to the source and libraries are permitted to
Electronic Engineering Department, University of Tripoli, photocopy. AJCITA is not responsible for opinions presented
Libya, E-mail: a.ganoun@uot.edu.ly. in its publication, they represent the views of the individuals.
Omar A. Abu-Ella, Assistant Professor at Electrical and
Electronic Engineering Department, University of Misurata,
Libya, E-mail: omarabuella@eng.misuratau.edu.ly.
Mustafa A. Abuzaraida, Assistant Professor at Information
Technology Department, University of Misurata, Libya,
E-mail: abuzaraida@it.misuratau.edu.ly.

P UBLICATION S TAFF
Asmaeil A. Ahteebah
Almadar Journal for Communications, Information
Technology, and Applications
Vol. 03, No. 01, April 2016
Contents

Editorial Page: Technologies for Next Generations Networks


Abdulla A. Abouda, Page 1.

Performance Analysis of Double Orthogonal Space-Time Block Codes


Rajab Legnain, Ian Marsland and Roshdy Hafez, Page 2.

INTERNET as Dictionary for Image Compression


Elhusain Saad and Keigo Hirakawa, Page 6.

Using Bayesian Hypotheses Testing to Improve Radial Basis Function Neural Network Performance
Khaled Mohamed, Page 11.

Design Principles and Practical Implementation of an Islanding Preventing Algorithm Based on


IEC 61850-7-420
Mike Mekkanen, Erkki Antila and Reino Virrankoski, Page 15.

Feedforward and Feedback Control of Smart Material Actuators


Mohamed M. Edardar, Page 22.

An Overview of Mobile Indoor Coverage: A Case Study of a Local Mobile Operator


Maoloud Y. Dabab, Page 31.
1

EDITORIAL PAGE
Technologies for Next Generations Networks

virtual appliances from different vendors. Operators need to be


able to mix and match hardware, hypervisors, and appliances
from different vendors without incurring significant integration
costs. Performance issues such as latency, reliability and
security are other issues that has to be addressed.
In spite of these challenges several cases has been reported
where operators started to partially utilize NFV and SDN,
however, significant research in the area of NFV and SDN
is still needed before complete benefits of these technologies
are realized.
Abdulla A. Abouda1 This is the third issue of Almadar journal for communi-

T ODAY’S networks are primary built with equipments cations, information technology and applications where six
from multiple vendors. Each equipment has a specific papers appear. The papers cover wide scope of topics, ranging
function and role in the network. Examples of these equip- from deep technical research papers to overview studies.
ments in mobile operator’s context are mobile switching cen- Utilizing spatial domain to increase data rate and improve
tre, home location register, short message server, charging and signal quality is very important area for research in wireless
billing equipments, gate ways, etc. When network operators communications. It has been widely known that using multiple
plan to lunch new services through these networks or to do transmit and receive antennas can increase channel capacity
network expansion, new equipments should be added and significantly. In this issue one of the articles presents an
spaces to accommodate them are required beside the need approximate closed-form expression for the double orthogonal
for energy and professional staff to run and integrate these space-time block code.
new equipments. Adding new equipment to an existing net- Image processing has several important applications in our
work consumes valuable time and introduces extra costs and daily life, however, in order to deal with images compromise
overheads. Also introducing changes to an existing network between image quality and size is always needed. Image
is a hard task to manage. Virtualizing network functions and compression for different purposes is also an important topic.
defining network connectivity using software are new concepts One of the articles in this issue proposes utilizing INTERNET
to overcome these challenges. as dictionary for image processing.
Tomorrow’s networks are built with equipment from multi- Neural networks have found it’s way in several applications,
ple vendors and commercial off-the-shelf hardware that is used and communications is one of these applications. One of
to create a generic platform and cloud based environment. In- the articles in this issue uses Bayesian hypotheses testing
stead of performing network functions using specific purpose to improve the performance of radial basis function neural
hardware, a software running on the generic platform is used to network.
implement the network functions. With this approach operators Comparison of different control schemes for smart material
can gain several benefits. Among these benefits are reduction actuators is presented in one of the articles.
in equipment costs and power consumption, increasing speed Communication technology has found it’s way in several
of time to market, improvement of network availability and areas. One of the articles in this issue presents design principle
scalability. Flexibility is another benefit that can be gained and practical implementation of one of these application in
when software defined networking (SDN) technology is in- power system.
troduced. These new technologies will introduce fundamental Since most of mobile traffic is generated inside buildings,
changes to the current network architecture and play essential indoor coverage is a very important issue for mobile system
role in the next generation networks. operators. One of the articles in this issue presents an overview
Despite the above mentioned benefits operators can gain of mobile indoor coverage using a case study from Almadar
from NFV and SDN, there are several challenges that has Aljadid Co.
to be addressed. One of these challenges is the coexistence Finally, in behalf of the editorial staff I would like to use
of legacy hardware based network system with the newly this chance to acknowledge the contribution of all people to
introduced NFV and SDN systems. Development in OSS/BSS this issue of the journal from authors, editors, reviewers and
has to go in line with NFV and SDN. Another challenge publishing staff. I also would like to invite researchers to send
that network operators will face is how to integrate multiple their contribution to the journal for publication.

1 Abdulla A. Abouda, Manager of Research and Development Office,


Almadar Aljadid Co. Libya, E-mail: a.abouda@almadar.ly.
2 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

Performance Analysis of Double Orthogonal


Space-Time Block Codes
Rajab Legnain1 , Ian Marsland1 and Roshdy Hafez1

Abstract—In this paper we study the performance of the double Notation: Throughout the paper, the following notations are
orthogonal space-time block code (DOSTBC) scheme. In the used. Bold lowercase and bold uppercase letters denote vectors
DOSTBC, the transmitter simultaneously transmits its symbols and matrices, respectively; [·]∗ , [·]T , [·]H and k·kF denote the
using two OSTBC systems. In this paper, we consider the case
where the receiver uses minimum mean square error detection complex conjugate, transpose, Hermitian, and Frobenius norm
to estimate the transmitted symbols. We derive an approximate of a vector or a matrix, respectively; E[·] denote the expecta-
closed-form expression for the bit error rate of the DOSTBC tion operation; In is used to denote the n × n identity
 matrix;
for the case of a slow uncorrelated Rayleigh fading channel with Q (x) denote the Gaussian Q-function; and nk = k!(n−k)! n!
is
perfect channel estimation at the receivers. This expression is 
valid for the well-known modulation schemes, (i.e., M -PAM, M - the binomial coefficient, where nk = 0 for k > n.
QAM and M -PSK). The analytical results are confirmed using
Monte Carlo simulations. II. S YSTEM M ODEL
Index Terms—Double orthogonal space time block codes, BER We consider a MIMO system with NT = 4 transmit
analysis. antennas and NR receive antennas, as shown in Fig. 1. The
transmitter splits an incoming 4 × 1 symbol vector, x, into
I. I NTRODUCTION two 2 × 1 vectors, and then it uses two OSTBCs to encode

O RTHOGONAL space-time block code (OSTBC) trans- these two vectors in space and time, and then transmits them
mission is a multiple-input multiple-output (MIMO) simultaneously over two time slots.
technique which is used to provide transmit diversity to
mitigate the channel fading. The main advantage of an OSTBC 1
is that it provides full spatial diversity with a simple detector. x1
OSTBC1 1
However, an OSTBC can only provide a spatial code rate that 2

Channel
x .. x̂
is equal to or less than one. For example, an OSTBC with S/P . Detector
two transmit antennas (i.e., Alamouti’s OSTBC [1]) can offer x2
3

a spatial code rate of one while OSTBCs with three or four OSTBC2
4
NR
transmit antennas can offer a maximum spatial code rate of
3/4 [2]. Transmitter Receiver
The spatial code rate can be doubled with sacrificing some Fig. 1: System model of DOSTBC.
transmit diversity, by using two OSTBCs. This scheme, called
double orthogonal space-time block code (DOSTBC) [3], The received sample vector at the receiver can be expressed
requires that the number of receive antennas should be equal as
to or larger than twice the spatial code rate [4]. For example,    
hk,1,1 hk,1,2 n1,1
DOSTBC with two Alamouti codes requires at least two
r  h∗k,1,2 −h∗k,1,1    n∗1,2 
receive antennas to detect the transmitted symbols. 2
X E    
 .. ..  xk,1  .. 
In this paper, we study the performance of the DOSTBC y=  . .  + . ,
NT   xk,2  
scheme with two Alamouti Codes, where the receiver uses k=1  hk,NR ,1 hk,NR ,2 | {z }  nNR ,1 
a minimum mean square error (MMSE) detector. The main h∗k,NR ,2 −h∗k,NR ,1
xk
n∗NR ,2
contribution of this work is the analytical evaluation of the | {z } | {z }
Hk n
bit error rate (BER) for DOSTBC scheme. More specifically, (1)
we derive an approximate closed-form expression for the BER, where E is the average energy transmitted per time slot. hk,i,j
when the transmitter either uses Gray coded M -PAM, M -PSK is the channel coefficient between the j th transmit antenna
or square M -QAM (i.e., M = 4, 16, 64). of k th OSTBC and the ith receive antenna, and is modeled
The rest of this paper is organized as follows. The system as independent and identically distributed (iid) zero-mean
model of the DOSTBC is described in Section II. In Section III complex Gaussian (ZMCG) random variables with variance
we derive the approximate closed-form expression of the BER T
σ 2 . xk = [xk,1 , xk,2 ] is the transmitted symbol vector from
for the scheme. Simulation results and the conclusions are the k OSTBC, and xk,v , v ∈ {1, 2}, is selected from a
th
presented in Sections IV and V, respectively. set of M complex symbols, i.e., xk,v ∈ {s1 , · · · sM }. The
1 The authors are with Department of SCE, Carleton University, Ottawa, covariance matrix of xk is given by Rxk = E[xk xH k ] = I2 . n
Canada, E-mails: {rlegnain, ianm, hafez}@sce.carleton.ca. is the effective noise vector, where the element ni,t represents
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 3

the noise on the ith receive antenna in the tth time slot, which where ρ12 = | NTEσ2 hHl,1 hl,2 | = 0 because hl,1 and hl,2 are
2
N
is modeled as a ZMCG random variable with variance σN2 . orthogonal. Therefore, (6) can be rewritten as
The receiver uses either a maximum likelihood (ML), mini- (
2
E 2
mum mean square error (MMSE) or zero forcing detector (ZF) kHl kF + σN , n = 1, 2
λl,n = N2T . (9)
to detect transmitted symbol vectors. In this paper we consider σN , n = 3, · · · , 2NR
the MMSE detector. The MMSE detector estimates the trans-
Using (5) we can express (4) as
mitted symbol vector from the k th OSTBC by multiplying the
received vector by a weight matrix that is given by E H −1
!−1 γk = u Λ uk,v , (10)
r 2 NT k,v l
E H X E
Wk = H Hl HlH + σN 2
I2NR . (2) where the vector uk,v = VlH hk,v = [uk,v,1 , · · · , uk,v,2NR ]T
NT k NT
l=1 has the same probability density function of hk,v , because Vl
The estimated symbol vector is given by x̂k = Dec (Wk y), is a unitary matrix (i.e., Vl VlH = I). We can rewrite (10) as
where Dec(·) denotes the hard decision operation and depends L
X E
on the transmitter signal constellation. Then the receiver com- γk =
2
|uk,v,n | , (11)
bines the estimated vectors to obtain the detected transmitted n=1
NT λ l,n
symbol vector, x̂ = [x̂1 , x̂2 ]T .
where L = 2NR is the number of independent channels
III. P ERFORMANCE A NALYSIS between the k th OSTBC and the receiver.
The average SER of detecting a symbol xk,v is found by
In this section we derive an approximate closed-form ex-
averaging (3) over the pdf of γk (or equivalently over the pdfs
pression of the average BER for the DOSTBC scheme. The
of λl,1 and uk,v,n ). Because it is difficult to take the average
following analysis is valid for any modulation schemes where
over the pdf of λl,1 , we use the average value of λl,1 , which
the instantaneous symbol error rate (SER) at an instantaneous
is given by
signal-to-noise (SNR) of γ can be approximated by [5]
p  L
λ̄l,1 = E [λl,1 ] = Eσ 2 + σN
2
. (12)
P s (γ) ≈ α Q 2βγ , (3) NT l
where α and β depend on the modulation scheme. For M - Therefore, the average SER of detecting a transmitted symbol
PAM, α = 2(1 − M 1
) and β = (M 23−1) , for M -PSK, α = 2 can be expressed as [7]
Z π/2  
and β = sin2 ( M
π
), and for square QAM, α = 4(1 − √1M ) and α 2β
P̄s ≈ Φγ k − dθ, (13)
β = 2(M −1) . Note that (3) is exact for BPSK.
3
π 0 2 sin2 (θ)
The instantaneous signal-to-interference-plus-noise ratio
(SINR) of xk,v at the detector output is where Φγk (ω) is the moment generating function (MGF) of
γk and is given by [7]
E H !Ns
γk = h M−1 hk,v (4)  L−Ns
NT k,v l 1 1
Φγk (ω) = , (14)
where hk,v is the v th column of the effective channel matrix, 1 − ωγ̄k 1 − ω λ̄ γ̄/σ
k
2
l,1 N
Hk , and Ml = NET Hl HlH + σN 2
I2NR (where l = 1, 2 and 2
σk E
l 6= k) is the interference-pulse-noise covariance matrix. Using where γ̄k = NT σN
is the average SNR per channel per symbol.
2

spectral decomposition we can rewrite Ml as The scalar Ns is the number of interfering symbols that are
Ml = Vl Λl VlH (5) transmitted from the lth OSTBC and is equal to two (i.e.,
Ns = 2), because we consider Alamouti OSTBC in which
where Λl is a diagonal matrix of the eigenvalues of Ml , and two symbols are simultaneously transmitted in each time slot.
Vl is the eigenvector matrix. The eigenvalues of Ml are [6] equation (13) can be rewritten as
 γl,1 +γl,2
 2
σN (1 + 2 + 12 Υ), n = 1
γ +γ Z π/2 2 
Y mn
λl,n = σN 2
(1 + l,1 2 l,2 − 12 Υ), n = 2 , (6) α sin2 (θ)

 2 P̄s ≈ dθ, (15)
σN , n = 3, · · · , 2NR π 0 n=1
sin2 (θ) − ψn
2
where γl,v = NTEσ2 hH E
l,v hl,v = NT σN2 kHl kF is the instanta- where ψ1 = βγ̄k , ψ2 = λ̄ βγ̄/σk
2 , m1 = L − Ns and m2 = Ns .
N l,1
neous SNR of the symbol xl,v , which has a chi-square distri- N
Using the results of the limit integral in [7], an approximate
bution with eight degrees of freedom, which has a probability closed-form expression for P̄s is
density function (pdf) of   NX
  α ψ 1 s −1

1 γkL−1 γk P̄s ≈ 
ψ2
L−1 Bm Im (ψ2 )
fγk (γk ) = exp − , (7)
(L − 1)! γ̄kL γ̄k 2 1− ψ 1
ψ2
m=0

and  ψ  L−N X s −1 
q −
1
Cm Im (ψ1 ) (16)
Υ= (γl,1 − γl,2 )2 + 4ρ12 (8) ψ2 m=0
4 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

100
where
 m
ψ2 Am
Bm = −1 L−1
, (17)
ψ1 m 10−1

 m NXs −1 m

ψ1
Cm = 1− n
L−1
 An , (18)
ψ2 n=0 n
10−2

BER
Ns −1
 Ns
Ns −1+m
Y
Am = (−1) m
(L − n) , (19) 10−3
(Ns − 1)! n=1
n6=m+1
Simulation: 256-QAM
Simulation: 16-QAM
and Simulation: 32-PSK
s 10−4 Simulation: 8-PSK
m  
ψ X 2n −n Simulation: 4-PSK
Im (ψ) = 1 − [4 (1 + ψ)] . (20) Simulation: BPSK
1 + ψ n=0 n Theoretical: solid lines
10−5
Finally, the BER is given by 0 5 10 15 20 25 30 35 40
γ̄ (dB)
P̄s
P̄b ≈ . (21) Fig. 2: BER performance of DOSTBC for MR = 2.
log2 M
IV. S IMULATION R ESULTS
100
In this section, we provide simulation results for the BER Simulation: 256-QAM
Simulation: 16-QAM
performance of the DOSTBC scheme using Monte Carlo Simulation: 32-PSK
simulation. These results are compared to the theoretical 10−1
Simulation: 8-PSK
Simulation: 4-PSK
results. We also compare the performance of the DOSTBC Simulation: BPSK
scheme with well-known schemes. Theoretical: solid lines
Fig. 2 shows the theoretical and the simulation results of 10−2
the average BER performance of the DOSTBC scheme for
BER

BPSK, 4-PSK, 8-PSK, 32-PSK, 16-QAM and 256-QAM. We


consider the case where NR = 2. As expected, the theoretical 10−3
results accurately match the simulation results at high SNR.
For higher values of M , at low SNR, there is an error between
the theoretical and simulation results due principally to the 10−4
approximation in (21). However, this error decreases as SNR
increases.
In Fig. 3, we extend the simulation to the case when NR = 10−5
0 5 10 15 20 25 30 35
4. Again, we show that the theoretical results approaches to
γ̄ (dB)
the simulation results as the SNR increases.
In Fig. 4 and 5, we compare the BER performance of the Fig. 3: BER performance of DOSTBC for MR = 4.
DOSTBC with the BER performance of vertical Bell Labs
layered space-time (VBLAST) and spatial modulation-spatial
multiplexing (SM-SMux) [8] schemes. We consider that the V. C ONCLUSIONS
VBLAST uses minimum mean square error ordered successive
interference cancellation detector [9], and the SM-SMux uses In this paper we studied the performance of the DOSTBC
the suboptimal detector [8]. In Fig. 4, we consider the case scheme. We derived an approximate closed-form expression
where all the schemes use only two receive antennas (i.e., of the BER for Gray coded M -PAM, M -QAM and M -PSK
NR = 2). We plot the BER performance of the 4 bps/Hz modulation schemes. In the analysis, we considered a slow
spectral efficiency for the 4 × 2 4-QAM DOSTBC, 2 × 2 uncorrelated Rayleigh Fading channel with perfect channel
BPSK VBLAST, and 4 × 2 4-QAM SM-SMux. In Fig. 5, estimation at the receiver. Using Monte Carlo simulation, we
we extend the comparison to the case where all the schemes showed that the theoretical results have excellent agreement
use four receive antennas (i.e., NR = 4). We plot the BER with the simulation results.
performance of the 4 bps/Hz spectral efficiency for the 4 × 4
4-QAM DOSTBC, 2 × 4 4-QAM VBLAST, 4 × 4 BPSK
R EFERENCES
VBLAST, and 4 × 4 4-QAM SM-SMux. As shown in Fig. 4
and 5, the DOSTBC scheme outperforms the VBLAST and [1] S. M. Alamouti, “A simple transmit diversity technique for wireless
SM-SMux schemes in both cases. This can be explained by communications,” IEEE Journal on Selected Areas in Commun., vol. 16,
the fact the that the DOSTBC can provide transmit and receive pp. 1451–1458, Oct. 1998.
[2] V. Tarokh, H. Jafarkhani, and A. Calderbank, “Space-time block coding
diversity while the other two schemes can provide only receive for wireless communications: Performance results,” IEEE Journal on
diversity. Selected Areas in Commun., vol. 17, pp. 451–460, March 1999.
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 5

100
DOSTBC: 4x2, 4-QAM
VBLAST: 2x2, 4-QAM
SM-SMux: 4x2, Na =2, BPSK

10−1
BER

10−2

10−3

10−4
0 2 4 6 8 10 12 14 16 18 20 22 24 26
γ̄ (dB)

Fig. 4: BER performance for a spectral efficiency of 4 bps/Hz and NR = 2.

100
DOSTBC: 4x4, 4-QAM
VBLAST: 2x4, 4-QAM
VBLAST: 4x4, BPSK
SMMUX:4x4, Na =2, BPSK
10−1

10−2
BER

10−3

10−4

10−5
0 2 4 6 8 10 12 14 16 18
γ̄ (dB)

Fig. 5: BER performance for a spectral efficiency of 4 bps/Hz and NR = 4.

[3] E. N. Onggosanusi, A. G. Dabak, and T. M. Schmidl, “High rate space-


time block coded scheme: Performance and improvement in correlated
fading channels,” in IEEE Wireless Communications and Networking
Conference, WCNC, Orlando, FL, USA, March 2002, pp. 194–199.
[4] R. Legnain, R. Hafez, and I. Marsland, “Multiuser detection for rate-
1/2 ostbc wih four transmit antennas,” in IEEE International Conference
on Communications, Signal Processing, and their Applications, ICCSPA,
Sharjah, UAE, Feb 2013.
[5] J. G. Proakis and M. Salehi, Digital Communications, 5th ed. McGraw-
Hill, 2008.
[6] J. Cui, D. Falconer, and A. Sheikh, “Analysis of BER for optimum
combining with two co-channel interferers and maximal ratio combining
with arbitrary number of interferers,” in Proc. of 7th IEEE PIMRC, Taipei,
Taiwan, Oct. 1996, pp. 53–57.
[7] M. K. Simon and M. S. Alouini, Digital Communication over Fading
Channels. Wiley, 2005.
[8] R. Legnain, R. Hafez, I. Marsland, and A. Legnain, “A novel spatial
modulation using MIMO spatial multiplexing,” in IEEE International
Conference on Communications, Signal Processing, and their Applica-
tions, ICCSPA, Sharjah, UAE, Feb 2013.
[9] R. Bohnke, D. Wubben, V. Kuhn, and K.-D. Kammeyer, “Reduced
complexity MMSE detection for BLAST architectures,” in IEEE Global
Telecommunications Conference, 2003. GLOBECOM’03., San francisco,
CA, USA, Dec. 2003, pp. 2258–2262.
6 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

INTERNET as Dictionary for Image Compression


Elhusain Saad1 and Keigo Hirakawa2

Abstract—The success of modern image compression algorithm based on some sparsity criteria [5], we exploit similarities of
depends on the effectiveness of the dictionaries that represent structures that occur naturally in images by searching over a
the images indirectly. Recent emphasis has been on the sparse large set of blocks or image patches that closely approximates
representation, where the redundant dictionaries are used to
closely approximate the image vector with a combination of a an image patch taken from the image to be compressed. To
few coefficients. In this work, we explore the possibility of using accomplish this, we make use of readily available images
INTERNET as a source of dictionary for image compression. on INTERNET as a “dictionary” that we search over, and
As the number of images available online are extraordinarily address them with the Universal Resource Locater (URL) that
large, the quality of image compression improves by searching uniquely identify the relevant dictionary images. Compression
for the closest dictionary matches. A unique requirements to this
work is the efficient addressing of these images by compressing is performed in part by employing text-based lossless com-
the Universal Resource Locater (URL). The results show that pression algorithms to reduce the bits required to specify the
the proposed algorithm is competitive with state-of-art image URL. When the search range is sufficiently large, the proposed
compression algorithms. compression algorithm attains state-of-art performance.
Index Terms—Dictionary learning, image compression. The proposed compression architecture is shown in Fig. 1.
As shown by this illustration, the encoder and decoder require
access to the INTERNET to retrieve the dictionary images.
I. I NTRODUCTION Such a structure makes sense if the storage cost is higher

P ROGRESS in storage and transmission technologies fu- than the cost of network access, or if the goal is required to
eled the advancements of multimedia applications like compress a large set of images at once (e.g. image services
graphics and video. Compression plays a critical role by such as Flickr, Instagram, Google Images, etc. may benefit
reducing the bits required to represent the same image or from the proposed approach). This approach also presents a
video data. Image compression exploits correlation between set of unique challenges, such as the permanence of URL
the neighboring pixels to represent redundant information that and the hashing of dictionary images to make the search
can be described with fewer bits [1]–[3]. Image compression more efficient. Though these questions remain, the potential
can be divided into lossy and lossless compression, where of image compression as evidenced by the qualities we attain
the latter results in exact reconstruction of the data while the cannot be ignored.
former would sacrifice image quality to reduce bits further.
When the quality loss is acceptable, the additional savings II. INTERNET AS D ICTIONARY
in encoded bits may be significant. As such, the focus of A. Criteria for Establishing A Match
research is on the favorable trade offs between compression
ratio and image quality loss. A common approach to lossy Let f ∈ RW ×W be the target block or image patch of size
image compression is based on some dictionary or a set of W × W cropped from image-to-be-compressed F : Z2 → R.
basis vectors designed to represent the image with only a few Let fˆ = αg + β be a candidate dictionary block that approx-
coefficients. For instance, the discrete cosine transform and imate f , where g ∈ RW ×W and α and β are parameters.
wavelet transform play key roles in JPEG and JPEG2000 com- The criteria we employ to determine the “goodness” of match
pression standards, respectively [1]–[3]. More recent develop- between f and fˆ are mean squared error (MSE) and structural
ments in wavelet-based compression methods include Embed- similarity index metric (SSIM) which predicts their perceived
ded Zero Tree Wavelet (EZW) which compresses images by similarity [6].
thresholding wavelet coefficients in a tree structure [1], Set MSE is defined as follows
Partitioning In Hierarchical Trees (SPIHT) which leverages W
X −1 W
X −1

inter-subband redundancies [1], [2], and Adaptively Scanned M SE(f, fˆ) =W −2 (f (m, n) − fˆ(m, n))2 . (1)
Wavelet Difference Reduction (ASWDR) that predicts active m=0 n=0

wavelet coefficients [4]. By taking derivative with respect to α and β to minimize MSE,
In this work, we employed a strategy based on the notion we arrive at the standard linear regression coefficients
that expanding the size of dictionary would improve the σf g σf σf g
overall coding efficiency. Instead of learning new dictionaries α̂ = 2 = ρf g , β̂ = µf − 2 µg . (2)
σg σg σg
1 Elhusain Saad is with Department of Electrical, University of Misurata, where µf and µg are the means of blocks f and g, respectively;
Misurata, Libya, E-mail: saadelhusain1@eng.misuratau.edu.ly, The author σf2 and σg2 are variances of f and g, respectively; and σf g and
thanks Libyan ministry of higher education for funding this research.
2 Keigo Hirakawa is with University of Dayton, Dayton, OH 45469 USA, ρf g = σf g /σf σg are covariance and the Pearson correlation
E-mail: khirakawa1@udayton.edu. coefficient, respectively. Consequently, substituting α̂ and β̂
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 7

Fig. 1: Block diagram of basic image compression framework (IAD1).

into equation (1) yields the minimum value of mean square corresponding to ρf ĝk̂ = max(ρf ĝ1 , . . . , ρf ĝK ).
error: The number of available dictionary images in INTERNET
is virtually infinite. To give an idea for how many dictionary
e2min = σf2 (1 − ρ2f g ). (3)
images are required to meet an “acceptable” image quality
The SSIM is defined by the relation: SSIM (f, f˜) = (say M SE = σf2 (1 − ρ̂2 ) or SSIM = ρ̂), let us treat f and
! ! ! Gk as i.i.d. random variables for the moment. The probability
2µf µf˜ + C1 2σf σf˜ + C2 σf f˜ + C3
, (4) that we would have met the desired quality with K dictionary
µ2f + µ2f˜ + C1 σf2 + σf2˜ + C2 σf σf˜ + C3 images is
where µf˜ and σf˜ are mean and variance of f˜; σf f˜ denotes n o n oK
P r ρf ĝk̂ > ρ̂ = 1 − P r ρf ĝ ≤ ρ̂ . (9)
covariance of f and f˜; and C1 , C2 , C3 are small constants. In
the special case that we use f˜ = α̃g + β̃ where In other words, one can expect the success probability to
σf σf increase as K grows (although the encoder complexity as well
α̃ = , β̃ = µf − µg . (5)
σg σg as the number of URL codes to retain would also increase).
instead of α̂ and β̂ in (2), (4) simplifies to (when C1 , C2 , C3
are negligibly small) III. P ROPOSED I MAGE C OMPRESSION A LGORITHM
A. Basic IAD Image Compression Framework
SSIM (f, f˜) = ρf g . (6)
We first describe the basic image compression architecture
B. Template Matching using INTERNET as dictionary (IAD), illustrated by Fig. 1.
Based on (3) and (6), we conclude that g that maximizes the Let the target image F : Z2 → R be partitioned into blocks
Pearson correlation coefficient ρf g ∈ [−1, 1] is the minimizer fij ∈ RW ×W , as follows: ∀(m, n) ∈ {0, . . . , W − 1}2 ,
of MSE error emin (using α̂ and β̂) and the maximizer of fij (m, n) = F (m + iW, n + jW ). (10)
SSIM (using α̃ and β̃). Given a large dictionary image G :
Z2 → R, we are interested in identifying an W ×W dictionary Based on the template matching of (7) and the maximum
block ĝ inside G that maximizes ρf ĝ : Pearson correlation criteria of (8), lossy storage/transmission
of fij (m, n) is accomplished indirectly via fˆij (for MSE-
ĝ = arg max ρf g (7) optimal compression) or f˜ij (for SSIM-optimal compression),
g(m,n)=G(m+∆m,n+∆n)
where
Known as template matching, this search task can be per-
formed very efficiently by combining a few convolution oper- fˆij = α̂ij ĝk̂ij + β̂ij , f˜ij = α̃ij ĝk̂ij + β̃ij . (11)
ators.
Suppose for the moment that the dictionary images Gk : As such, the following quantities must be stored/transmitted
Z2 → R are drawn at random from INTERNET, where k by the encoder in order for the decoder to be able to recover
denotes the image index k ∈ {1, 2, . . . }. Let ĝk = G(m + fˆij or f˜ij :
∆mk , n + ∆nk ) be the best candidate dictionary block in • µfij and σfij (or σfij · ρfij ĝk for fˆij )
dictionary image Gk corresponding to the Pearson correlation • (∆mi , ∆ni ) for locating gk̂ in Gk̂
coefficient ρf ĝk and the location (∆mk , ∆nk ). Obviously, we • URL for locating image Gk̂
choose as the final candidate as the maximizer of ρf ĝk over The URLs are stored in a “text file” in ASCII format, each
k ∈ {1, 2, . . . , K} dictionary images: separated by carriage returns—this text file will be com-
fˆ = α̂ĝk̂ + β̂ pressed subsequently using standard text-based compression
subject to k̂ = arg max ρf ĝk (8) techniques. The remaining values were stored in a separate
f˜ = α̃ĝk̂ + β̃ k
“data file” after being quantized into 8bit integers (i.e. 4 × 8
8 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

Fig. 2: Block diagram of multilevel image compression framework (IAD2).

(a) PSNR vs bpp (b) SSIM vs bpp (c) Effects of varying K (d) Effects of varying W 0

Fig. 3: The performance of the proposed algorithm and some standard methods. (K=number of dictionary image, W 0 size of block of block means).

bits per W × W block). Contrast this to the fact that original B. Simple Extensions
W × W block taken from an 8bit image requires 8W 2 bits to
represent. There are a few simple extensions that one can employ
to improve the overall image quality. First, block based
The exact steps taken by the encoder are as follows:
compression methods (such as JPEG) are subject to blocking
1) Retrieve K dictionary images {G1 , . . . , GK } found on- artifacts, where the discontinuities at the block boundaries
line (random draws from Flickr.com in our implemen- become visible. Traditionally, blocking artifacts are dealt with
tation) and cache them to local storage. in post-processing by smearing pixels near the boundaries. In
2) For each target block fij , locate the template matched the context of IAD compression, however, the decoder may
block ĝk that maximizes Pearson correlation coefficient impose “feathering” by w pixels. That is, by cropping an
ρfij ,ĝk in each dictionary image Gk . (W +2w)×(W +2w) image patch from Gk̂ instead of W ×W
3) Identify the dictionary block ĝk̂ among {ĝ1 , . . . , ĝK } image patches, we allow for an overlap between neighboring
that maximizes Pearson correlation coefficient. blocks by w pixels that can be blended together for smoother
4) Output µfij , σfij , (∆mi , ∆ni ) to “data file” and URL transition. In our investigation, we found that feathering was
of Gk̂ to “text file.” far more successful at removing blocking artifacts than the
5) Once all target blocks are encoded, then apply an effi- standard post-processing.
cient lossless compression to “data file” and “text file” Second, recall (9). Although the success probability would
(the default Zip compression in our implementation). increase as K grows, this also requires more template match-
ing. One simple technique to improve the success probability
The exact steps taken by the decoder are as follows: is to modify (8) as follows:
1) Decompress “data file” and “text file.”
2) Retrieve dictionary images from online by parsing URLs k̂ = arg max |ρf ĝk |. (12)
k
in “text file.”
3) For each reconstruction of target block fˆij (or f˜ij ), That is, when ρf ĝk is close to -1, fˆ (or f˜) reconstructed
locate the template matched block ĝk̂ (m, n) = Gk̂ (m + with negative α̂ (or α̃) is actually very good approximation
∆mi , n + ∆ni ). Compute µg and σg . to f . Allowing α̂ to be a signed variable would essentially
4) Compute α̂ and β̂ (or α̃ and β̃), and reconstruct fˆij (or mean the number of dictionary images is doubled to 2K,
f˜ij ). therefore increasing the success probability by a significant
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 9

(a) Original (b) IAD1 (c) IAD2


BPP=8, SSIM=1, PSNR=∞ BPP=.33, SSIM=.95, PSNR=28.45 BPP=.27, SSIM=.94, PSNR=28.32

(d) JPEG2000 (e) ASWDR (f) SPIHT


BPP=.33, SSIM=.95, PSNR=33.6 BPP=.48, SSIM=.93, PSNR=31.7 BPP=.30, SSIM=.92, PSNR=30.8
Fig. 4: Reconstructed cameraman image. (BPP=bits per pixel).

margin compared to (9): decoding of Encoder 1 bitstream. This risk can be mitigated
n o n o2K by increasing the search range K for Encoders 2 and 3, as well
P r |ρf ĝk̂ | > ρ̂ = 1 − P r ρf ĝ ≤ ρ̂ . (13) by reducing the size of “blocks of block means” and “blocks
of block standard deviations” to W 0 × W 0 where W 0  W .
However, this technique has the disadvantage that the signed
α̂ variable requires an extra sign bit to encode.
IV. E XPERIMENTAL R ESULTS
The proposed algorithms was implemented using Matlab.
C. Multilevel IAD Image Compression 64 INTERNET dictionary images have been downloaded from
The proposed IAD image compression can be applied Flickr.com with varied texture, feathers, and size. Color dic-
recursively. Consider Fig. 2. Here, Encoder 1 is the same tionary images were converted to grayscale images. We chose
encoder found in Fig. 1, with same input/output structure. to compress some standard grayscale reference images (cam-
Encoder 2 is also identical to Encoder 1, except for the fact eraman, Peppers, Boat, etc.). The performance of the single
that the input image is replaced by the “block means” µfij . In level (IAD1) and multilevel (IAD2) compression algorithms
other words, the block means when taken together as a whole were compared to other well known compression methods
can be interpreted as another image, which can be compressed (SPIHT, ASWDR, JPEG2000). Fig. 3 reports the PSNR and
by Encoder 2 in the same manner that Encoder 1 does (for SSIM scores averaged over the compressed images. Based on
instance, “blocks” in Encoder 2 are really “blocks of block Fig. 3(a-b), IAD1 and IAD2 are comparable to the alternatives
means”). Similarly, Encoder 3 is responsible for compressing in PSNR performance, but clear winner in SSIM performance
the “block standard deviations” σfij . The decoder will have for bits per pixel ranges between 0.02 and 1. Fig. 3(c) shows
to process the streams from Encoders 2 and 3 to recover µfij that expanding the dictionary image size would indeed result in
and σfij first before the actual image f can be decoded from improved image quality, while Fig. 3(d) shows the sensitivity
Decoder 1. of multilevel IAD compression algorithm (IAD2) to block
The multilevel IAD compression results in a higher com- sizes (W 0 ).
pression rate by reducing the number of bits required to code Example images in Fig. 4 show surprising contrast between
µfij and σfij . However, the image quality suffers also because IADs and the alternative methods compressed at similar rates.
of the fact that error in µfij and σfij can propagate to the IADs retain the sharpness of image edge better and the texture
10 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

in the sky are preserved. One might argue that loss of grass
textures occurs more gracefully in IADs than SPIHT, ASWDR,
and JPEG2000.

V. C ONCLUSION AND F UTURE W ORK


We proposed a new compression algorithm based on “IN-
TERNET as dictionary of images” principles to attain accept-
able images at high compression ratios. Match between target
image block and dictionary block was made based on MSE and
SSIM optimal criteria, and the stored/transmitted bitstreams
included text files to record the URL of the dictionary images.
The experiments showed that we are competitive in PSNR and
surperior in SSIM performance when compared to SPIHT,
ASWDR, and JPEG2000. Future work in this are include
hashing of dictionary image blocks to enable a fast encoding
over a very large set of dictionary images.

R EFERENCES
[1] R. Sudhakar, R. Karthiga, and S. Jayaraman, “Image compression
using coding of wavelet coefficients–a survey,” International Journal on
Graphics, Vision and Image Processing, vol. 5, no. 6, pp. 25–38, 2005.
[2] A. Said and W.A. Pearlman, “A new, fast, and efficient image codec
based on set partitioning in hierarchical trees,” Circuits and systems for
video technology, IEEE Transactions on, vol. 6, no. 3, pp. 243–250, 1996.
[3] P.D. Saraf, D. Sisodia, A. Sinhal, and N. Gupta, “Comparisons of
wavelets based image compression methods,” World Journal of Science
and Technology, vol. 2, no. 3, 2012.
[4] J.S. Walker and T.Q. Nguyen, “Adaptive scanning methods for wavelet
difference reduction in lossy image compression,” in Image Processing,
2000. Proceedings. 2000 International Conference on. IEEE, 2000, vol. 3,
pp. 182–185.
[5] R. Rubinstein, T. Faktor, and M. Elad, “K-svd dictionary-learning for
the analysis sparse model,” in Acoustics, Speech and Signal Processing
(ICASSP), 2012 IEEE International Conference on. IEEE, 2012, pp.
5405–5408.
[6] Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli, “Image
quality assessment: From error visibility to structural similarity,” Image
Processing, IEEE Transactions on, vol. 13, no. 4, pp. 600–612, 2004.
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 11

Using Bayesian Hypotheses Testing to Improve


Radial Basis Function Neural Network Performance
Khaled Mohamed1

Abstract—Radial basis function (RBF) neural network is one based on Bayesian hypotheses testing. In general, a Bayesian
of the networks which are used to classify different types of hypothesis test is a method of making decisions using data.
patterns. Using RBF is the key behind the success of this network. These data can be either from a controlled experiment or
In this network, the detector (conventional detector) in training
and testing data usually depends on the maximum value of the an observational study . In statistics, we call a statistically
output or a certain threshold, however it doesn’t always assure significant if it is unlikely to have occurred by chance alone,
the chosen threshold can give the best performance. In this paper, according to the significance level, and a pre-determined
Bayesian hypothesis testing is introduced and applied to RBF threshold probability. Critical tests of this kind may be called
neural network to get the optimum threshold. We replace the tests of significance, and when such tests are available we may
conventional detector in the training and testing data by the
Bayesian detector (smart detector). A comparison between the discover whether a second sample is or is not significantly
two detectors is conducted which shows that the smart detector different from the first [8]. The smart detector depends on the
can dramatically decrease the average cost of miss and false probability density function when an event occurs. Bayesian
alarms and lead to less error percentage. hypotheses testing are the principle of our smart detector
Index Terms—Radial basis functions, Clustering and pattern, where the average risk can be minimized. The smart detector
neural network, Bayesian hypothesis, smart detector. can minimize the error percentage by accessing to the training
data and getting the probability density functions and applying
these functions to the testing images. Moreover, changing the
I. I NTRODUCTION
miss and false alarm costs can lead to changing the system

R ADIAL basis function (RBF) networks are usually


utilized for approximating functions from given in-
put–output patterns. The performance of a trained RBF net-
performance so choosing the right costs can improve the
system performance.
The rest of this paper is organized as follows. The the-
work depends on several factors. The locations and number oretical description is defined and discussed in Section II.
of the RBF, their shape and the method used for learning In Section III, we demonstrate the algorithm efficacy using
the input-output mapping are some important factors [1]. RBF handwritten numbers database. Finally, in Section IV, we offer
network consists of a two-layer feed-forward neural network conclusions.
such a network is characterized by a set of inputs and a
set of outputs. In between the inputs and the outputs layer,
there are hidden layers. Each of them implemented a radial II. T HEORATICAL D ESCRIPTION
basis function. In the pattern classification application, the
network input usually represents features while the output A. Radial Basis Function Neural Network
corresponds to a pattern or a class. A pattern which has
same features is classified in a same group. Then weights Radial functions consist of a two-layer feed-forward neural
multiplied to those groups where a certain output is gotten network such a network is characterized by a set of inputs and
so the detector can decide what the input is. The conventional a set of outputs. In between the inputs and the outputs layer,
detector usually gives one to the output whose value is bigger there are hidden layers. Each of them implemented a radial
than a certain threshold; however it doesn’t always assure the basis function. The network input usually represents features
chosen threshold can give the best performance. while the output corresponds to a class in the pattern classifi-
RBF networks were firstly introduced into the neural net- cation application. The mean for each class are multiplied by
work application in 1988 [4]. Since then many methods have the weights where the M th outputs from the weights block
been proposed to improve RBF networks performance and will be inputs to the detector. The detector decides which the
utilize to many applications [2], [5]–[7]. Many methods have input is (either depends on a certain threshold or the highest
focused exclusively to optimize the number of the RBF, their value of the output). Fig. 1 shows a simple block diagram for
shape and the method used for learning the input-output training a system whose patterns have N elements, L hidden
mapping [3]. layers and M outputs.
In this paper, we focus in how the threshold can be RBF function are used to classify different patterns into L
optimized. We take a completely different detector by viewing groups where the outputs of RBF block are computed as
the design of Bayesian detector (smart detector). This detector 1
uk = 2 (1)
1 Khaled Mohamed Ahmied Mohamed is with the Department of Electrical 1 + e(x−µk ) /2σk
and Computer Engineering, University of Al Jabal Al Gharbi University,
E-mail: alkmeshe@yahoo.com. where k = 1, 2, ..., L and σk is the standard deviation. Then
12 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

XP u1 V1
1

XP u2 V2
2

RBF function Weights Detector Y


(L × N ) Wij (M × L) (M × 1)
(the detected output)

XP uL VM
N

Classification all patterns Weights which multplied The detector is to detect


into L groups by classified groups what the input is

Fig. 1: Shows simple block diagram for training a system for RBF network.

the output of weights block is computed in such a way as C. Bayesian detector (Smart detector)
L The smart detector depends on the probability density
X
Vj = Wji ∗ ui , (2) function when H0 occurs as well as the probability density
i=1 function when H1 occurs. In order to implement this type of
detector, we have to follow the next steps:
where j = 1, ..., M (M is the number of the output from the
• First, train the system with the conventional detector.
weights block) , i = 1, ..., L (L is the number of hidden layers
• Then, we replace the detector with the smart one and use
or clusters) . The last step to get the output is computed such
as  the same training images again as inputs and also keep
+1, Vj ≥ C the trained weights to get the probability density function
yj = (3)
0, Vj < C for each digit.
• Use two dimension kernel density function to smooth
where C is a certain threshold value. In the conventional
the probability density function (PDF) for P (y|H0 ) and
detector, the threshold value is usually chosen to be the
P (y|H1 ) where y in our case is combinations of V1 and
maximum value in the vector Vj .
V2 .
• Compute the threshold τ by assign costs for all Cij and
B. Bayesian Hypotheses testing apply them to Equation (6).
Let assume a simple case which has only two possible hy- • Apply the smart detector in the testing part so by checking
potheses H0 and H1 (M = 2), correspondence to two possible the likelihood ratio whether is bigger than the threshold
probability distribution P (H0 ) and P (H1 ) respectively where: value (τ ) or not so the smart detector can decide what
H0 : Y ∼ P (H0 ) versus H1 : Y ∼ P (H1 ) [8]. the input is.
The Bayesian decision rule is given by
 III. R ESULTS AND D ISCUSSION
1 if L (y) ≥ τ
δB (y) = (4)
0 if L (y) < τ In this section, fully connected network structure of 784
input nodes (N = 784), 2 output nodes (M = 2), and 30
where the likelihood ratio is
hidden nodes (L = 30) are implemented by Matlab. Our
P (y/H1 ) database is a set of (28×28) pixels handwritten numbers
L (y) = , (5)
P (y/H0 ) images which consists about 6000 ones and 6000 zeros where
the threshold value is half of them are used in the system training and the other
half are used in the system testing so it is obvious that the
P (H 0 ) . (C00 − C10 )
τ= , (6) priors are equal (P (H0 ) = P (H1 ) = 0.5). In order to compare
P (H1 ) . (C11 − C01 ) fairly between the two approaches, learning factor parameter
and Cij is the cost when my decision is i while Hj is true. and number of hidden nodes are fixed where the learning
The Bayesian hypotheses testing minimize the average risk parameter is set to 0.1 and number of hidden nodes is set to
by using 30. The fact we have chosen these numbers because it gives
the least error percentage when the conventional detector is
rB (δ) = P (H0 ) [C10 P (Γ1|H0) + C00 P (Γ0 |H0 )] +
(7) applied. The metric which is used in the comparison between
P (H1 ) [C11 P (Γ1 |H1 ) + C01 P (Γ0 |H1 )] the conventional and the smart detectors is the error percentage
where Γ1 and Γ0 are defined in Equations (8) and (9) for each digit. Using Equation (6), the threshold value is 3
respectively (τ = 3) when C00 = C11 = 0, C10 = 3 and C01 = 1.
Γ1 = {y ∈ Γ|L (y) ≥ τ } (8) The inputs of the detector (V1 and V2 ) have different values
when all the input images have zero labels as well as when
all images have one’s labels. Fig. 2 shows the values of V1
Γ0 = {y ∈ Γ|L (y) < τ } (9) and V2 when H0 and H1 occur. It is clear from Fig. 2 that
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 13

2
H0 is occured
1.5 H1 is occured
1
1
0.8
0.5

P(V1,V2 /H1)
0.6
V2

0
0.4
-0.5
0.2
-1
0
-1.5 2
1 2
-2 0 1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 0
V1 -1 -1
V2 -2 -2
Fig. 2: Values of V1 versus V2 when H0 and H1 occur. V1

Fig. 4: The probability density function for the likelihood when H1 occurs.

the images whose label are ones have different locations from
images whose labels are zeros except some small overlapping
between them. Those overlapping should be distinguished in Equation (4).
order to decrease the error percentage for each digit. This We run the experiment 50 times using the conventional and
distribution of data is very important to get the probability smart detector where the mean error percentage for each digit
density functions. In order to get the probability density is computed and reported. In each run, randomly a set of 3000
function for the likelihood of H0 , around 3000 images whose ones and a set of 3000 zeros are picked and utilized for training
labels are zeros are used as inputs into the training system. while the rest of images are used for testing.
Fig. 3 shows the probability density function for likelihood
Fig. 5 shows the mean error percentage for zero and one
when H0 occurs (P (V1 , V2 /H0 )). Likewise the probability
when the conventional detector is used while Fig. 6 shows
density function for the likelihood of H1 can be gotten, around
the mean error percentage for zero and one when the smart
3000 images whose labels are ones are used as inputs into the
detector is used. Table (I) shows the average error percentage
training system. Fig. 4 shows the probability density function
comparison between the conventional and smart detectors for
for the likelihood when H1 occurs (P (V1 , V2 /H1 )).
each digit. It is obvious from the experiment that there is a
big improvement in terms of the error percentage when the
conventional detector is replaced by the smart one. The im-
provement factor is 10 so the system performance is improved
1 ten times when the smart detector is conducted. This idea can
0.8
be generalized with more than two outputs (not only H0 , H1 )
by only knowing the probability density functions for those
P(V1,V2 /H0)

0.6 inputs.
0.4
Error Percentage for the digits zero and one using the conventional detecor
2
0.2 Hidden nodes is 30 and learning parameter is 0.1
1.8
0
2 1.6
1 2
1 1.4
0
Error Percentage %

0
-1 -1 1.2
V2 -2 -2
V1 1

Fig. 3: The probability density function for the likelihood when H0 occurs. 0.8

0.6
The probability density function for likelihood when H0 or
H1 occur is corresponding to the dense of V1 and V2 so the 0.4

probability density functions have big values where there is 0.2

a dense in data. The likelihood ratio can be computed by 0


knowing the values of V1 and V2 and plug them into the 0
numbers
1

probability density function for H0 and H1 as in Equation (5)


Fig. 5: Error Percentage for zero and one when the conventional detector was
where y represents by V1 and V2 . Then, the decision rule can used.
decide whether the input is zero or one which depends on
the likelihood ratio and the threshold value as illustrated in
14 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

Error Percentage for the digits zero and one using the smart detecor
2
Hidden nodes is 30 and learning parameter is 0.1 [8] Vicent Poor ,”an Introduction to Signal Detection and Estimation”,”
1.8 Springer Science and Business Media”, 1995.
1.6

1.4
Error Percentage %

1.2

0.8

0.6

0.4

0.2

0
0 1
numbers

Fig. 6: Error Percentage for zero and one when the conventional detector was
used.

TABLE I: S HOWS THE AVERAGE E RROR P ERCENTAGE C OMPARISON B E -


TWEEN THE C ONVENTIONAL AND S MART D ETECTORS FOR EACH D IGIT.

Average
The digit 0 1
Error
percentage %
1.61% 1.4% 1.5%
(Conventional Detector)
Error percentage %
0.2% 0.091% 0.15%
(Smart Detector)

IV. C ONCLUSION
Bayesian hypotheses testing are the principle of our smart
detector where the average risk can be minimized. Bayesian
detector can minimize the error percentage by accessing to
the training data and getting the probability density functions
and applying these functions to the testing images. In addition,
changing the miss and false alarm costs can lead to changing
the system performance so choosing the right costs can deliver
a higher performance. In average, the smart detector enhances
our system performance by ten times comparing to the con-
ventional detector. Finally, smart detector can play a key rule
for enhancement many systems whenever there is available
data for training.

R EFERENCES
[1] Mark J. L. Orr, “Introduction to Radial Basis Function Networks,” center
for cognitive science, 1996.
[2] R. Assaf, S. El Assad, Y Harkouss, and M. Zoaeter,”Efficient classifi-
cation algorithm and a new training mode for the adaptive radial basis
function neural network equaliser”, ”Communications, IET”, Vol. 06,
No. 02, pp. 125−137, 2012.
[3] Song, Shenmin, Zhigang Yu, and Xinglin Chen,”A novel radial basis
function Neural network for approximation”, ”International Journal of
Information Technology 11”, No. 09, pp. 246−257, 2005.
[4] D. S. Broomhead, and D. Lowe,”Multivariable Functional Interpolation
and Adaptive Networks”, ”Complex Systems”, No. 02, pp. 321−355,
1988.
[5] Luo Huaien and S. Puthusserypady, “Bayesian radial basis function
network for modeling fMRI data,” in Engineering in Medicine and
Biology Society, Vol. 01, No. 01, pp. 450−453, Sep. 2004.
[6] D. Fisch, and B. Sick,”Training of radial basis function classifiers with
resilient propagation and variational Bayesian inferenc”, ”IJCNN 2009.
International Joint Conference on Neural Networks”, pp. 838−847,
2009.
[7] C. Andrieu, N. Freitas, and B. Aand ,”Robust Full Bayesian Learning
for Radial Basis Networks”, Neural Computation, Vol. 13, No. 10,
pp. 2359−2407, Oct. 2001.
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 15

Design Principles and Practical Implementation of


an Islanding Preventing Algorithm Based on IEC
61850-7-420
Mike Mekkanen1 , Erkki Antila2 and Reino Virrankoski3

Abstract—Incoming booming of Distributed Energy Resources Islanding, term represents the case where, a segment(s) from
(DER) and the growing complexity of electrical systems are the electrical grid is isolated from the main Utilities energy
driving the need for robust communication solutions to refurbish supply due to the maintenance schedule or a fault that may
the existing automation and protection functions or even invents
a new functions. In order to sustain an efficient electrical energy comprise at least one DER which might continue to energize
systems operation and reduce hazard for the Electrical Utilities the islanded segment independently.
persons within such dynamic and adaptable systems structure
behavior that handle numerous parameters and variables. Design This case and from the electrical grid managements point
and practical implementation of an islanding detection and view [3] creates unacceptable situation in terms of power
preventing technique is carried out. The designed approach relies
on the electrical grid communication system network based on quality that need to be maintained within the consumer
the IEC 61850-7-420 standard. Electrical grid distributed main assigned acceptable range [4], [5]. Moreover, it may cause
management service (DMMS) takes the response to monitor all a severe damage to the DERs in case of reconnection to
the electrical connection points (ECP) and the DERs status. The electrical grid while losing synchronism with the rest of the
developed solution has overcomes the lack of the existing remote electrical grid [6]. Lastly, it provides a hazard in case of
(transfer-trip) and local (active, passive) detection techniques,
since, upon its implementation, measurements for any electrical restoration or fault clearing to the electrical grid persons who
grids operation parameters (e.g. frequency, active and reactive involved with the repairing tasks. The main Utility supply
power, etc.) are not needed. The designed DMMS and data objects and the DERs are planned to operate in parallel through the
modeling for the predefined LNs within the IEC 61850-7-420 and electrical grid rather in the islanding mode. Therefore, in terms
their implementations through a specifically designated testing of the electrical grid managements assigned requirements
is applied within the commercial instruments. The obtaining
results show that the correctness for the design of the islanding for the DERs that aimed to connect to the main electrical
protection function through the DMMS approach and the data network must install islanding protection on its intertie to the
modeling of various IEC 61850-7-420 predefined data object LNs. electrical network. Energize segment(s) from the electrical
grids by DERs are not allowed in case of isolation from the
Index Terms—DERs, IEC 61850-7-420, Islanding detection, main electrical grid. As a result, anti-islanding gained high
Smart grid. attention from both industry [7] and academia, and several
anti-Islanding detection techniques have been proposed.
These proposed anti-islanding techniques can be divided
I. I NTRODUCTION
into two categories local technique and remote technique.
HE expected increase of the electrical energy
T consumption to double by 2050 based on the International
Energy Agency study and the incoming boom of the DERs
Local technique can be divided into active technique and
passive technique. All these above mentioned techniques have
several advantages and disadvantages that might limit their
are the most global challenges facing the energy Utilities [1]. implementations [8].
Decentralized DERs lead to parallel operation within the Within the local technique category, and firstly the active
electrical distribution network (Utilities’ main generators techniques, the main idea is to inject a small disturbances
and DERs) which involves several challenges. The most to the electrical network, real time monitoring to the DERs
challenging is the detecting of the islanding scenarios, or loss outputs need to be considered. These small disturbances are
of main. Since, the existing distribution networks were not expected to make significant changes within the outputs of
intended to handle the incoming boom of the DERs scattered the DERs in islanding case. Whereas, the DERs outputs
along to the distribution electrical grid. Moreover, they are changes are almost inconsiderable in connection within the
designed upon the passive network assumption [2]. main electrical grid cases. The main advantage of the active
technique is the ability to detect the islanding case upon
1 Mike Mekkanen is with University of Vaasa, Department of
the perfect matching between the load demands and the
Computer Science, Communications and Systems Engineering, E-mail:
mike.mekkanen@uva.fi. DERs output power. However, several disadvantages are
2 Erkki Antila is with University of Vaasa, Department of Electrical En-
associated with the active technique implementation, such
gineering and Energy Technology, P.O. Box 700, FI-65101 Vaasa, Finland, as, introducing disturbances within the electrical grid might
E-mail: erkki.antila@uva.fi.
3 Reino Virrankoski is with University of Vaasa, Department of Computer lead to effect on overall energy system quality. Further, in
Science, Communications and Systems Engineering, E-mail: rvir@uva.fi. case of more than one DER within the islanded segment,
16 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

interference between the introduced disturbances and DERs way. Since the IEC 61850-7-420 offers a set of extended LNs
output signals may affect the implementation of the active that support several DER units, more details within the rest of
techniques. In addition, the long implementation time upon this paper [19].
creating the external disturbances and monitor the DER
output parameters might reduce its effectiveness. Several
II. DER S S TANDARDIZATIONS B ROUGH BY IEC
advanced islanding detection techniques within the active
61850-7-420
techniques have been used to detect the loss of main situation
(e.g., reactive power error detection, impedance measurements Utilities and DERs manufacturers announce the concept of
technique, phase shift or frequency technique and hybrid growing need to define and standardize the communication
techniques that monitor hybrid parameters). However, most outside the individual SAS that may include various DERs
of the active techniques share the same characteristics, higher IEDs. As a result, the standard IEC 61850-7-420 was
cost, complexity, introduces disturbances to the electrical published in 2009 as an extension of the IEC 61850
network, long time to implements etc. that might limits its standard and to address these issues. The IEC 61850-7-420
implementation and rarely been used within the electrical specify various types of LNs and information modeling
network [9]. that applicable for the various DERs e.g. fuel cell systems,
photo-voltaic (PV), combined heat and power (CHP), etc [19].
The other techniques within local category is the passive The predefined information modeling, LNs facilitates the
techniques, which overcome most of the active limitation for communication and the integration of the DERs into the
instance easy to use, no need to introduce disturbances to Utilities protection and automation systems. Utilities and DER
the electrical network, less cost etc. Several passive islanding manufacturers are expected to achieve benefits from utilizing
detection techniques have been proposed, for instance rate of the IEC 61850-7-420 in terms of reducing the installation,
change of frequency (ROCOF) [10], over/under voltage [11], maintenance costs. Further, offering standardization of all
[12], phase displacement [13], change of the DERs power DERs data models that will improve the interoperability
outputs [14], and hybrid solution [15]. All the proposed among distributed automation system (DAS) and DERs and
passive techniques share the same concept that monitoring increase the reliability of the energy grid [20].
and measuring the outputs of the DERs parameters through
the interface points to the main electrical grids. However, IEC 91850-7-420 standard address the information
indiscriminate between the normal electrical grid transient modelling for various DERs, LNs whereas, other IEC 61850
events such as (DERs out from supplying power, instantaneous implementation aspects such as the services modeling,
loads changes, faults in some electrical grids unites that leads assigned system configuration language (SCL) and the
to out from operation etc.) and the actual islanding scenarios mapping schemes over the defined protocols had been
may cause a nuisance tripping that reduces the reliability covered within the previous IEC 61850 10 parts first
and availability of the electric grids. Moreover, uncertainties version. These DERs information modeling LNs defined
in islanding detection upon the measuring outputs of the by IEC 61850-7-420 standard involve not only for the
DERs parameters for instance (ROCOF), when the local loads local communication among the local DERs and the local
are perfectly matches with the local DERs generated power management service systems, however, they may support
(dead zone). Therefore, and form the above discussion these the sharing information with the main grids operators or
islanding detection techniques might not function as intended, aggregators who manage the whole electrical grid operation.
and reduces the reliability and availability of the electrical The defined DERs LNs based on IEC 61850-7-420 have been
grids upon false tripping or increase the hazard to the electrical grouped into four groups upon their operation characteristic
grids Utilities persons [16]–[18]. (node classes and common data classes (CDC)). These DERs
With these motivations, in order to increase the reliability LNs groups are logical nodes for DER management systems,
and availability of the electrical grids, reduces the hazard to logical nodes for DER generation systems, logical nodes
the electrical grids Utilities persons and improve the remote for specific types of DER and logical nodes for auxiliary
techniques, since the existing transfer-tripping and power line systems. These predefined DERs LNs represents all the DERs
carrier signaling (PLC) lack of co-ordination between the DER operation aspects parameters such as for instance, connecting
units and the main electrical grids. According to the existing status, availability status, economic dispatch parameters,
remote techniques, there are no specific real communication start/stop time, operating mode etc. however, according to
standards and protocols for electrical systems that have the the islanding detection case study within this paper, two
ability to provide to their customers for monitoring, controlling DER LNs have been considered the DPST and the DRCS
and protection the DERs devices. As a result, a developed LNs. The DPST provides the real-time ECPs status and
islanding detection algorithm upon the the standardized com- measurements,(the ECPs are usually associated with each
munication solution brought by IEC 61850-7-420 DERs LNs DER, load, lines Buses etc. that need to connect to the local
and IEC 61850 communication system has been developed. power system, group of DERs that need to interconnect to
This developed algorithm has the ability to monitor all the the Utility energy system i.e.). Where, the DRCS LN defined
electrical connection points (ECPs) and DERs status within the control status of individual DER or group of the same
the electrical grid, and share the information, take the right type of the DERs that controlled within individual controller.
decisions to handle the electrical grid operation in efficient
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 17

Monitor all the (ECPConn)


III. I SLANDING D ETECTION S CHEME AND C ASE S TUDY DPST LNs for the all ECPs

Performance evaluation of any developed function it may


Monitor all (ModOnConn) DRCS LNs for
require within the first step to define the test scenario that the DERs and Main supply status

supposed to be representing aspects almost related to realty


operation conditions. According to the developed islanding
detection scheme two main operation aspects need to be moni- No Is any of the (ECPConn)
DPST LNs is False
tored, that the status of the all ECPs and the status of the DERs
within the electrical grid. This task can be achieved through
monitoring the DERs LNs that specified by the IEC 61850-7- Yes
Update electrical grid
420 standard for instance the common data object (ECPConn) topology

for the DPSTs LNs take the values True or False that represent
the status of the electrical grids units whether they are elec-
trically connected at the ECPs or not. Whereas, the common No
Is the LNs of the main
(ECPConn) DPST is False or the
data object (ModOnConn) and (ModOffAval) for the DRCSs (ModOffAval) DRCS is True

LNs both take the value True that represents the status of the
Yes
DERs when they are ON and connected electrically within the Is any of the DERs LNs
(ECPConn) DPST are True but not
No Yes
grid or when they are OFF but available to start respectively. connected to the main and the Set the alarm condition
(ModOnConn) DRCS LNs
Where, all the available real time information that provided by are True

the IEC 61850-7-420 DERs LNs standard are gathered through Send signals to The active DERs to
change the values of the
the electrical grid communication network based on the IEC (ModOffConn) DRCSs LNs to True

61850 general object oriented substation events (GOOSE)


messages within the DMMS. In the DMMS, the available Send Trip signals to change the
values of all active DERs
gathered information can be used to detect and prevent the (ECPConn) DPSTS LNs to False

islanding scenarios. As a result, islanding detection algorithm


upon the gathered information from the various microgrid Fig. 1: Flow chart for the proposed islanding detection algorithm.
nodes has been proposed. The proposed islanding detection
algorithm facilitates the updating of the microgrid structure
and the indiscriminating between the normal electrical grid measurements lead to false tripping, in which that reduces
transient events and the actual islanding scenarios that increase the reliability and availability of the electrical power systems.
the reliability and availability of the electric grids as illustrated In addition, the proposed approach reduces the hazard for the
in Fig. 1 [22]. electrical systems persons, since no DERs or energy storages
First of all in order to better understand the new islanding still continue to supply segment(s) of the electrical micro
detection algorithm, lets firs considered a prototype electrical grids after the loss of main supply.
microgrid within various circumstances as shown in Fig. 2.

The prototype designed electrical microgrid consist of


a medium voltage main supply, transformer, four buses,
three various types of DERs, various types of loads, energy
storage, circuit breakers, lines, communication network based
on IEC 61850 standard and DMMS. Within the DMMS,
information is gathered from all the electrical microgrid units
through the communication network upon the IEC 61850
standard. It is worth noticing here that the new islanding
detection approach possess a flexibility through its operation,
since it is monitoring all the electrical microgrid units and
update its internal operation that related to the dynamic
behavior of the electrical microgrid and takes the decisions
upon the final electrical microgrids topologies. As a result,
the new islanding detection approach overcome the lack of
the existing remote technique (inter-tripping) that it can be
implemented regardless to the electrical network topologies
“plug and play” scheme. Furthermore, the developed islanding
detection approach has overcome the lack of the existing
local (active, passive) detection techniques, since, upon its Fig. 2: Prototyped microgrid designed concept.
implementation, measurements for any electrical parameters
such as frequency, active and reactive power etc., are not Considering the normal operation of the proposed prototype
needed. Where, uncertainties in these electrical parameters electrical microgrid is usually the circuit breaker CB7 is open.
18 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

In case that the CB8 is opened upon failed or maintenance among the microgrid system units, DERs and microgrid
cases, the DMMS detect the new electrical grid topology instruments need to be modeled in compliance with the
and attempts to closed CB7 to stabilized the electrical grid IEC 61850-7-420. In addition these models need to be also
operation and update its internal operation that related to equipped with full communication module which will receive
the new toplogy. In case that the CB7 out of operation the final islanding detection decision from the DMMS and
in any reason. As a result, islanding scenario has been publish their parameters such as the DERs status, ECP
occurred; since segment from the electrical grid (Bus 4) status to the DMMS and microgrid instrument subscribers
is not connected to the main and has it’s on local energy upon the control and protection functions implementations.
suppliers as in this case DER3, Energy storage 1. The second Fig. 3 illustrates the modeled DRCS and DPST LNs that
islanding scenario can be occurred when the CB3 and CB7 represent the generic DERs and the electrical connection
are opened, causing that unplanned new electrical microgird points respectively within the microgrid simulation.
topology, segment form the electrical microgrid (Bus 2) is
islanded upon the local DER1. The third islanding scenario
From Fig. 3 the DRCS and DPST LNs consist of several
can be occurred when the CB4 and CB8 are opened, causing
data object parameters which are listed as “Mandatory” or
that segment form the electrical grid (Bus 3) is islanded upon
“Optinal”. According to the DRCS LN modeling through
the local DER2. The forth islanding scenario can be occurred
the proposed islanding detection algorithm three of the
when the CB4 and CB7 are opened, causing that segments
mandatory DRCS LN data object parameters are used to
form the electrical grid (Bus 3, Bus 4) are islanded upon the
represent the various DERs controller status. These data
local DER2, DER3 and energy storage 1. The fifth islanding
object parameters are DRCS.ECPConn, DRCS.ModOnConn
scenario can be occurred when the CB3 and CB8 are opened,
and DRCS.ModOffAval. The DRCS.ECPConn indicates that
causing that segments form the electrical grid (Bus 2, Bus 4)
the DERs are electrically connected to the ECP when this data
are islanded upon the local DER1, DER3 and energy storage
parameter is set to “True”. Where, the DRCS.ModOnConn
1. Lastly, islanding scenario can be occurred when the CB3
indicates whether the DER is in operation mode “ON” and
and CB4 or CB2 are opened, causing that segments form the
electrically connected when this data parameter is set to
electrical grid (Bus 2, Bus 3 and Bus 4 for the first case and
“True”. Lastly, the DRCS.ModOffAval indicates whether
all Buss within the second case respectively) are islanded
the DER is “OFF” mode and available to start. Whereas,
upon the local DER1, DER2, DER3 and energy storage 1.
the DPST LN through the DPST.ECPConn data object
All the announced expected islanding scenarios can be easily
parameter that indicates the connection of the microgrid units
detected within the new islanding detection techniques and
at ECP when this data parameter is set to “True”. Control
prevented, since the DMMS attempts to real time monitoring
status of various microgrid DERs data objects LN can be
all the ECPs and DERs status through the DERs LNs upon
extracted from the above mentioned three DRCS data object
the IEC 61850-7-420 standard and within the high speed
parameters. The extracted DERs control status including with
electrical grid communication network based on the IEC
the DPST ECP data object parameter are need to be gathered
61850 standard. However, in order to increase the reliability
by the DMMS through the communication system network
of the electrical microgrid and reduces hazard the existing
for proper implementation of the proposed islanding detection
islanding detection techniques should be used as backup for
function.
the new islanding proposed approach in case of failures of
the communication network.
Within the DMMS modeled block final islanding detection
decision making functions are designed upon the gathered
information from various microgrid nodes. Where, two
IV. DATA O BJECT M ODELING WITH IEC 61850-7-420
islanding detection decision making functions are designed
S TANDARD
that represent two different scenarios. The first islanding
In this section virtualization for the proposed prototyped detection decision making function (scenario one_Function1)
microgrid system units through data object modeling using is monitoring the main supply status and the ECP to the
the LNs that predefined by the IEC 61850-7-420 international electrical grid that will be simulated and explained in depth
standard has been made. Communication among the various within the next section. Whereas, the second islanding
microgrid system units LNs and DMMS through the IEC detection decision making function (scenario two_Function2)
61850 communication protocol general object oriented is monitoring the entire microgrid ECPs (various islanding
substation events (GOOSE) using a communication system cases are explained in section III) and the DERs status in
network. Communication system network is the essential for order to define and prevent the islanding scenarios both
the realization of the proposed islanding detection algorithm functions are illustrated in Fig. 4. Various simulation cases
scheme. Matlab/Simulink simulation software has been used are presented in previous work [23]. It is worth noticing here
through the microgrid system units LNs, DMMS modeling that scenario two will not simulate since the idea for islanding
and commissioning the entire microgrid system. detection is all most the same as in scenario one, where the
only differences are the inputs for the islanding detection
In order to achieve a proper implementation of the proposed decision making function as illustrated Fig. 4 Function 2.
islanding detection method and facilitate the communication
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 19

requires a validation [23]. In this section the design principles


ECPConn and practical implementation of the islanding detection and
ECPConn preventing algorithm based on IEC 61850-7-420 approach
DPST_block
has been carried out within a simple microgrid that consist of
one incoming feeder (The main power supply) and one out
ECPConn going feeder the include DER for simplicity as illustrated in
ECPConn Fig. 5. However, the same idea can be implemented for big
microgrids. In Fig. 5 the single line diagram (SLD) for the
AutoMen
proposed microgrid within the ABB substation automation
Terminator
builder 600 (SAB600) that the substation structure has been
Loc carried out. Where, the DMMS is used to make and publish
Terminator
the islanding detection decision upon the designed islanding
detection functions. A specifically designated testing is
ModOnConn
applied within scenario one that external signal virtual input
ModOnConn (VI_2) is generated to change the ECP status of the CB1
ModOnAval DPST_CB1.ECPConn data object parameter from True to
Terminator False for the incoming feeder within the DMMS function one.
ModOffAval

ModOffAval

ModOffUnav

Terminator

DRCS_block
Fig. 3: DPST and DRCS LNs blocks.

In_1

In_2
OR

In_3
Fig. 5: SLD within the SAB600 substation structure.

Whereas, the external signals (DI_1, DI_2 and VI_1) are


In_4 used to represent the Utility main power status upon the
LNs parameters DRCS.ECPConn, DRCS.ModOnConn and
Function 1
DRCS.ModOffAval since there is no commercial IEDs to
provide these parameters based on the IEC 61850-7420 stan-
The entire microgrid ECPs dard. Changing the ECP status of the CB1 through the
ststus upon DSPT
DPST_CB1.ECPConn LN data object parameter to “False”
AND
indicates that the main power supply is not any more elec-
The entire microgrid DERs trically connected to the main grid and supplies power, al-
ststus upon DRCS
though segment(s) from the electrical grid are still energized
Function 2 by the DER(s) (islanding scenario). This changing status
of the DPST_CB1.ECPConn LN data object parameter is
Fig. 4: Islanding detection decision making functions within DMMS. monitored through the DMMS islanding detection decision
making function1 Fig. 4. As a result DMMS publish a trip
signal to all the DERs whether to trip out from the electrical
V. P RACTICAL I SLANDING D ETECTION AND P REVENTING grid or turned off or both to prevent the islanding scenarios.
I MPLEMENTATION F UNCTION AND VARIABLES L INKING According to the DMMS COM600 configuration and within
the communication structure Logic_Processer_OPC_Server,
Substantiate the implementation and operation of the Logic_Processor_Subnetwork and Logic_Processor_IED need
designed functions and modeled LNs blocks through the to be added as illustrated in Fig. 6.
proposed protection scheme are required either within an From the Logic_Processor_IED and within the
experimental testing or a simulation. Simulation is very Logic_Editor window for the DMMS Function 1 need
attractive since it is cost-effective solution; however it may to be configured upon adding the required function block and
20 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

the microgrid within the ABB com600 gateway devise as


illustrated in Fig. 9. However, there is still some issues upon
the final configuration between the COM600 and different
manufacturers IEDS need to be solved.

Fig. 6: COM600 communication tree.

Fig. 8: Cross reference tool to link variables.

Fig. 7: DMMS function 1 configuration within the Logic Editor.

defined the input and output logic variables as illustrate in


Fig. 7. Those logic variables need to be linked to COM600
data objects within the Logic_Process_IED through the cross
reference tool as illustrated in Fig. 8. In addition the logic
variables can be linked directly to COM600 data objects as
before or they can be available as IEC 61850 data object
within the Logic_Processor _IED through the created virtual
LN for any other SAS implementations. From Fig. 8 the four
input variables for the DMMS function have been linked with
the LNs data objects within IED V52r which are digital input
(DI1 and DI2) Virtual input (VI1 and VI2). Where these LN Fig. 9: COM600 units’ connection status.
data objects simulate the IEC 61850-7-420 LNs (DPST and
DRCS) data objects. Where VI2 represent the signal based
on the main power supply circuit breaker positions whether
it is closed or open. If it is closed it send True signal (the VI. C ONCLUSION
main power supply connected to the grid) and based on the This paper has presented the design principles and practical
DMMS islanding function it inverts the input In_4. Therefore implementation of the islanding detection and preventing algo-
there is no trip signal roaming to the DER circuit breaker. rithm approach that involves the IEC 61850-7-420 LNs (DPST
Whereas, if the mean CB open it send False signal (the and DRCS) within monitoring the electrical microgrid units
main power supply not connected to the grid) and based status tasks. The DMMS with the internal islanding detection
on the DMMS islanding function it inverts the input In_4. function is used to gather the available information from the
Therefore, the DMMS function initiate a trip signal to the microgrid units, to detect and prevent the islanding scenarios.
DER CB to open. This trip signal that has been initiated Thanks to the IEC 61850 standard that offers a high speed
from the DMMS function need to be linked with the IED LN communication system network for the various manufacturers’
data object such as in our case VI2 within the IED V52 as electrical units. As a result, the islanding detection and pre-
illustrated in Fig. 8. Then the VI2 roaming signal has been venting practical implementation has been carried out in which
used in IED V52 to open the CB for the DER (trip out from that overcomes the lack of the existing islanding detection
the grid). Upon the above mentioned process that the DMMS techniques remote and local techniques, for instance, upon
gathered the information from the various LNs and executing the local techniques it has overcomes the lack of the existing
the islanding designed internal function and roaming the (active, passive) local detection techniques. Since, upon its
trip signal that preventing the islanding scenarios within implementation, measurements for any electrical parameters
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 21

such as frequency, active and reactive power etc., are not [17] Xu, W., & Mauch, K., “An Assessement of Distbuted Generation
needed, where, uncertainties in these electrical parameters Islanding Detection Method and Issues for Canada”, CETC-Varennes
2004.
measurements lead to false tripping, in which that reduces [18] Vivek, M., & Hashem, M., “A Hybrid Islanding Detection Technique
the reliability and availability of the electrical power systems. Using Voltage Unbalance and Frequency Set Point,” IEEE Transaction
In addition, it reduces the hazard for the electrical systems on power system, vol. 22, NO. 1, 2007.
[19] IEC 61850-7-420, “Communication Networks and Systems for Power
persons, since no DERs or energy storages are still continue Utility Automation for Distributed Energy Resources (DER)”, 2009.
to supply segments of the electrical grids after the loss of [20] T. Selim, C. Ozansoy, and A. Zayegh,"Simulation of Communication
main supply. The designed DMMS and data objects modeling Infrastructure of a Centralized Microgrid Protection System Based on
IEC 61850-7-420", IEEE Smart Grid Standard Symposium, vol. 27, No.
for the predefined LNs within the IEC 61850-7-420 and their 3, 2012.
implementations through a specifically designated testing is [21] Math Works, “Simulink Dynamic System Simulation for MATLAB,”
applied within the commercial instruments. The obtaining [Online], http://www.mathworks.com/help/pdf_doc/simulink/sl_gs.pdf.
[22] M. Mekkanen and E. Antila, “Islanding Detection Algorithm Using
results show that the correctness for the design of the islanding IEC 61850-7-420,” Protection Automation & Control World Conference
protection function through the DMMS approach and the data (pacworld 2015), Glasgow, UK, 2015.
modeling of various IEC 61850-7-420 predefined data object [23] [23] M. Mekkanen and E. Antila, Reino Virrankoski and Kimmo
Kauhaniemi “IEC 61850-7-420 Data Object Modeling for Smart Control
LNs. In addition, they show the successful implementations Islanding Detection”, IEEE WiSPNET conference, 2016.
of the designed islanding protection function and the mod-
eled microgrid units LNs upon the reliable detection of the
islanding scenarios.

R EFERENCES
[1] VTT Technical Research Center of Fin-
land (2009), “Energy Visions 2050,” [Online].
http://www.vtt.fi/files/vtt/energyvisions/visions2050_summary.pdf.
[2] T. Selim, C. Ozansoy and A. Zayegh, “Modeling of a Centralized Mi-
crogrid Protection System and Distributed Energy Resources According
to IEC 61850-7-420”, IEEE Transactions on power system, vol. 27, No.
3, 2012.
[3] FINGRID,“Annual-report”, (2013), [Online].
http://www.fingrid.fi/en/news/News
[4] N. Hatziagyriou, H. Asano, R. Iravani & C. Marnay, “Microgrids,” IEEE
power & energy magazine, vol. 5, no. 4,pp. 78-94, Abbrev. July-Aug,
2007.
[5] C. Colson & M. Nehrir, “A Review of Challenges to Real-Time Power
Management of Microgrids,” IEEE Power & Energy Society General
Meeting, vol. ED-11, no. 1, pp. 34–39, 2009.
[6] T. Mariano & L. Assis, "Automatic Reconnection From Intentional
Islanding Based on Remote Sensing of Voltage and Frequency Signals,"
IEEE Transaction on Smart Grid , vol. 3, no. 4, pp. 34-39, December,
2012.
[7] H. Laaksonen, D. Ishchenko & A. Oudalov, ”Adaptive Protection and
Microgrid Control Design for Hailuto Island,” IEEE Transaction on
Smart Grid, , vol. 5, no. 3, pp. 34-39, May, 2014.
[8] V. Menon M. Hashem, “A Hybrid Islanding Detection Technique Using
Voltage Unbalance and Frequency Set Point,” IEEE Transaction on
Power System. vol. 22, no. 1, pp. 34-39, February, 2007.
[9] Dysko, A., Burt, G., & Bugdal, R., “Novel Protection Methods for
Active Distribution Networks with High Penetrations of Distributed
Generation”, University of Strathclyde report 2006.
[10] X. Ding, P. A. Crossley & D. J. Morrow, “Islanding Detection for
Distributed Generation,” Journal of Electrical & Technology vol. 2, no.
1, pp. 19-28, 2007.
[11] C. Hsieh, J. Lin and S. Huang, “Enhancement of islanding-detection of
distributed generation system via wavelet transform-based approaches,”
Elsevier Journal of Electrical Power and Energy Systems, 2008.
[12] D. Mahinda P. Chiang, and Y. Li, “Protection of Microgrids During
Utility Voltage Sags,” IEEE Transaction on Industrial Electronics. vol.
22, no. 1, pp. 34-39, February, 2006.
[13] X. Ding and P. Crossly, "Islanding Detection for Distributed Generation,"
IEEE Power Tech Conference, pp. 1-4, 2005.
[14] A. Colet-Subirachs, ARuiz, Oriol Gomis, F. Alvarez, and A. Su-
dria,"Centralized and Distributed Active and Reactive Power Control
of a Utility Connected Microgrid Using IEC61850," IEEE SYSTEM
Journal. vol. 1, no. 1, March, 2012.
[15] H. Laaksonen, “New Multi-criteria-based Algorithm for Islanding De-
tection in Smart Grid,” IEEE PES Innovative Smart Grid Technologies
Europe, 2012.
[16] Kim, J. E., & Hwang, “Islanding detection method of distributed
generation units connected to power distribution system”, International
conference on power system technology, 2000.
22 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

Feedforward and Feedback Control of Smart


Material Actuators
Mohamed Edardar1

Abstract—Smart materials are used widely in many appli-


cations in nano-positing industry. The objective of this paper
is to compare different control schemes of these materials in
order to have perfect positioning. Although they have nice
features, the smart materials exhibit high nonlinearity, which
requires careful control strategies. The plant is represented as
a linear system preceded by a nonlinear operator. We discuss
both feedforward and feedback control methods used to improve
system tracking. The feedforward scheme relies on using an
approximate operator as a feedforward compensator. The design Fig. 1: A general control system for nonlinear actuators.
and analysis of three different feedback methods is presented
including the conventional Propertional-integral, sliding-mode,
and adaptive control. A comparison of the performance for these
methods is given in the simulation section. Some of these methods are using the feedforward principle
by inverting the nonlinearity of the system. This usually
I. I NTRODUCTION requires modeling the non-linearity in order to find the
inverse of the model and combine them to cancel it. Some
CTUATORS which used in nanopositioning nowadays
A are made of smart materials such as piezoelectrics [1],
magnetostrictives [2], and shape memory alloys [3]. These
good results for this scheme are reported in the literature [6],
[7]. Those methods have all the disadvantages of open-loop
system such as model imperfection and deviations due to
actuators have many desirable features such as high resolution, environmental effects. Feedback control methods are used
high stiffness, and fast frequency response [4]. However, they with considering any nonlinearity as a disturbance and
exhibit some nonlinear phenomena such as hysteresis, creep, assuming an upper-bound of the nonlinearity is known [8],
and resonance harmonics which badly affect the positioning [9]. Alternatively, combining the two controls as shown in
or tracking process [5]. In this paper, we will discuss the Fig. 1. reported to give better results which is very important
hysteresis and how to compensate for it. Some research is in systems require perfect tracking as in nano-positioning
dedicated for the other drawbacks. However, they do not have actuators [2]. Modeling of hysteresis and its inverse is the
major effects as the hysteresis and we will not consider them most tough task. Fortunately, many models are available
in this paper. with different accuracy and complexity. The most popular
Hysteresis is a complex nonlinearity which can be found in ones are Preisach operator [10], [11], Prandtl–Ishlinskii (PI)
many applications such as nanopositioning. This phenomenon operator [12], [13].
is seen in some materials by applying a quasi-static voltage Researchers also attempt different control methods to
on it. By increasing the applied input, then deceasing it we improve the system tracking. Conventional PID controller
see a loop which is known by hysteresis loop. If this loop is is examined with good results [14], [15]. More complicated
formed by applying the full range, it is called the major loop. robust [16]–[22] and adaptive methods [23]–[26] are also
By branching from any point on the major loop we obtain presented.
smaller side-loops known as minor-loops. One important In this paper, we compare different feedback methods under
characteristic about hysteresis that when the input decreases three main categories used in controlling systems comprising
to zero the output may display some reading. This property of hysteresis. Actually, these methods among others are explored
storing energy describes the hysteresis as a memory element. by researches and they are well-established in the literature.
All of these characteristics of the hysteresis make the task However, the comparison among these methods still requires
of controlling it hard. The obvious first task is to model a lot of work. For instance, a method succeeds to produce
the hysteresis. Some researches tried to describe hysteresis a good performance may fail if we have a different input
by differential equations but their models failed to capture trajectory. Sometimes, the tracking results presented by
all the hysteresis properties. Instead, different mathematical authors are excellent when the operating frequency is low
operators are used to describe them. for a certain method compared with others. But, it would not
outperform them if the frequency is increased. We start by
Control methods are suggested to improve the positioning showing how to model the hysteresis by different operators
by eliminating the nonlinear behavior of these actuators. and how to compensate it by using feedforward compensation.
1 Mohamed Mohamed Edardar is with the Department of Electrical Engi- Then, we present different feedback control methods which is
neering, University of Tripoli, Libya, E-mail: edardarm@gmail.com (M. E.). integrated with the feed-forward compensation in all of them.
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 23

40
exp. data

Magnitude, dB
plant model

20

0
500 5000
0
Phase, deg

−200

−400
500 5000
Frequency, Hz

Fig. 3: The relay operator with parameters α and β .


Fig. 2: Experimental results used to model the linear dynamics.

A proportional-integral (PI) control is a simple controller B. Nonlinear Operators


which gives good results at low frequency. A feed-forward Models of hysteresis have evolved from two different
branch is added to the PI controller to improve the tracking. branches of physics: ferromagnetism [27] and plasticity theory
We also present another robust method used frequently in [28]. The roots of both branches go back to the end of th 19th
nonlinear systems, which is known by Sliding-Mode-control century [6]. The core of this theory is formed by the so-called
(SMC). Then a Model Reference-Adaptive-control (MRAC) hysteresis operators which describe hysteretic transducers as
is chosen to compare an adaptive method with the former a mapping between function spaces. It is only in the 1970’s
robust method. when researchers started to couple the mathematical theory of
ordinary differential equation with hysteresis operators [29],
The remainder of the paper is organized as follows. In [30]. In the 1990s, engineers employed this theory on a
Section II, we presented the modelling of the Piezo-electric larger scale to develop modern strategies for the linearizion
actuator. This is followed by describing the feedforward of hysteretic nonlinearities with an inverse feedforward con-
compensation in Section III. In Section IV, we provided troller [6].
the design of three different feedback control approaches. The Preisach model [31]–[37] for electromagnetic hysteresis
Simulation results of these methods are presented in dates from 1935 [38]. This model uses a superposition of
Section V. Finally, we sum-up by concluding remarks in simple independent relay operators shown in Fig. 3. It has
Section VI. been successful in the modeling of hysteresis observed in
ferromagnetic, magentostrictive, and piezoelectric materials.
However, in designing with these smart materials, one has to
II. P IEZO - ELECTRICAL ACTUATOR M ODELING determine the density function for the preisach operator by
using input-output behavior of the material at hand [39]. This
It is common to model the Piezo-electrical actuator by
requires complex algorithms and computations to implement
linear plant followed or preceded by hysteresis. The hysteresis
it.
in this work is modeled by Prandtl–Ishlinskii (PI) operator
Recent papers also reference the so-called Prandtl-Ishlinskii
which is sub-class of Preisach operator. We will give a brief
(PI) operator [6], [13], [24], [40] which belongs to an im-
introduction to both of them in this section.
portant subclass of the Preisach operator [12]. The main
advantages of this approach are reduced model complexity of
A. Linear dynamics modeling the PI-operator in comparison with the Preisach operator and
The linear dynamics of the system for our example are the fact that the compensator of an invertible Prandtl-Ishlinskiii
identified as in Fig. 2. It is noted Fig. 2 that the plant has large operator can be calculated analytically. This allows an efficient
bandwidth and a second order system is a good approximation implementation of the compensator for real time application.
of the system. We used MatLab code to find the best fit for
these data which can be written in the state form as
C. Prandtl-Ishlinskii (PI) Operator
ẋ1 = x2
ẋ2 = −a1 x2 − a0 x1 + a0 (ud + d) (1) The PI operator dates to 1944 [38], and was proposed as
a model for plasticity-elasticity. The foundation of this model
where x1 = y denotes the position and x2 denotes the velocity is using elementary hysteresis operators, which have simple
of the positioner.. The parameters a0 and a1 are positive, and mathematical structures. One of the most familiar and most
their values are given in the simulation section. important elementary hysteretic mapping between the input
24 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

Fig. 4: The play operator with threshold paramter r. Fig. 5: Illustration of a hysteresis loop with piecewise linear characteristics.

signal V and the output U is so called play or backlash


operator, which is shown in Fig. 4 and described by the
following equation. The output of a PI operator [13] can be
represented as
Γ[V ] = wT Hr [V, z0 ](t) (2)
where, Hr = [Hr0 Hr1 . . . Hr,n−1 ]T is the vector of individual
play operators, wT = [w0 w1 . . . wn−1 ] is the vector of weights
Fig. 6: A system with hysteresis preceded by an inverse operator.
for individual play operators, rT = [r0 r1 . . . rn−1 ] with 0 =
r0 < r1 < . . . < rn−1 < ∞ represents the set of thresholds for
the play operators, zT0 = [z00 z01 . . . z0,n−1 ] denotes the initial
Krasnoselskii-Porkovskii (KP) operator [41] among others.
condition of the PI operator, and n is the number of play
operators.
Fig. 6 illustrates the system with a feedforward inverse
The play operator is defined by
hysteresis compensator. We assume that the actual hysteresis
Hri [V, z0 ] = max{V − ri , min{V + ri , z0i }} (3) is represented by an operator Γ p , defined by a vector of play
thresholds and a vector of play weights w∗ . We further assume
where i = 0, 1, . . . , n. that a nominal model Γm for the hysteresis is identified for
To simplify the notation, we will not explicitly put the initial implementation of Γ−1 m , an approximate inverse to Γ p . It is
condition for operators Hri , Γm , etc. assumed that Γm shares the same set of play thresholds with
Γ p , but its weight vector w differs from that of Γ p , which
represents the source of model uncertainty. In particular, we
III. F EEDFORWARD C ONTROL write w∗ = w + ∆w , with ∆w denoting the weight perturbation.
The control ud is applied to the inverse model and d = ud − u
The control methods that deal with hysteresis are classi-
is the inversion error.
fied into open-loop inverse compensation or feedback control
The uncertainty in the weight vector is translated into
methods, which usually include the inverse operator. Inverse
uncertainties in the slope and intercept of each segment i of
open-loop methods [6], [7], [13] demonstrate that they can
the hysteresis loop, denoted as ∆mi and ∆γi , and i = 1, 2, · · · , l,
reduce the tracking error appreciably, but they are susceptible
where l is the number of segments. Let the input-output
to model uncertainties and environmental changes.
relationship for the operator be
We calculate a bound on the inversion error by representing
the hysteresis nonlinearity as piecewise linear characteristics; u = m jv + γ j (4)
in other words, all hysteresis loops (major loops and minor
loops) consist of linear segments, where each segment si has and the input-output relationship for the inverse-operator be
a slope mi and an intercept γi with the output axis, and 1
i = 1, 2, · · · , l, where l is the number of segments. See Fig. 5 v= ud + γinv, j (5)
mj
for illustration. The bound on the inversion error is calculated
using the slopes and intercepts of linear segments. Note that By inserting (5) into (4) we obtain
this method is not confined to the PI operator; indeed, it applies 1
to wide class of models used in the literature, including the u =m j ( ud + γinv, j ) + γ j
mj
piecewise linear model adopted in [7], the Prandtl-Ishlinskii
(PI) operator [6], [24], the modified PI operator [13], and the = ud + m j γinv, j + γ j (6)
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 25

Fig. 8: Hysteresis inverse in the feedback path.

IV. F EEDBACK C ONTROL


Fig. 7: Inversion error bounds using slop-intercept method.
A general approach in coping with hysteresis is to
construct an inverse operator and integrate it with feedback
Hence, for perfect inversion we require techniques. An example of the nonlinear actuators is the
piezo-actuated positioner which is used to demonstrate our
m j γinv, j + γ j = 0 (7)
results in simulation section. These actuators are modeled as
Fig. 7 illustrates how these uncertainties can be used to an operators followed by linear dynamics. In this paper, The
determine an upper-bound on the inversion error. The term PI operator is used for demonstrating the results because its
u1 represents the output u when the inversion is perfect, and inverse can be calculated analytically.
its slope is one, while u2 represents the output in the presence
of uncertainties, The tracking error defined as
u1 (t) = ud (t) (8)
e = y − yr
∆m j where y = x1 is the measured position and yr is the reference
u2 (t) = (1 + )ud (t) + ∆dc, j (9)
mj trajectory.
where ∆dc, j can be obtained from (7) by including the uncer-
tainties as
∆dc, j =∆m j γinv, j + ∆γ j A. PI Controller
∆m j γ j − m j ∆γ j In this section we will briefly describe the components of the
= (10) closed-loop system as illustrated in Fig. 8. The linear dynamics
mj
of the plant are represented by a singularly perturbed system.
The difference between u2 and u1 represents the size of the
The bandwidth of the dynamics ωn is assumed to be large and
uncertainty and is denoted by d(t):
of the order 1/ε, where ε is a small positive parameter. The
d(t) =u2 (t) − u1 (t) model of the linear plant is given by
∆m j
=(1 + )ud (t) + ∆dc, j − ud (t) (11) ε ż = Az + Bu, y = Cz (16)
mj
The upper bound for each segment is where A is a Hurwitz matrix, B and C are matrices with
proper dimensions, and z is the state vector. We assume that
|γmax ||∆m,max | the feedback controller is a proportional-integral controller,
|∆dc,max | ≤ |∆γ,max | + (12)
mmin represented as
The upper bound for all segments is
ẋ = e = yr − y = yr −Cz, w = ki x + k p e (17)
∆mmax
|d(t)| ≤
|ud (t)| + |∆dc,max | (13)
mmin where e is the tracking error and w is the output of the
where mmin is the smallest slope, |∆mmax | is the largest slope proportional-integral controller. A feedforward path with a
uncertainty, and |∆dc,max | is the largest intercept uncertainty gain g is used to compensate for the DC gain of the linear
among segments and can be obtained from (10) as dynamics. When g = 0, the scheme falls back into the general
scheme that combines hysteresis inversion (in the feedback
|∆mmax γmax | + |mmax ∆γmax | loop) and feedback control. The DC gain of the plant is
∆dc,max ≤ (14)
mmin h = −CA−1 B.
A bound on the inversion error can be obtained from (13) as The input to the inverse operator, ud , and its output, v, are
expressed as
|d| ≤ k1 + k0 |ud | (15)
|∆mmax | 1
where ≤ k0 and |∆dc,max | ≤ k1 . ud = gyr + ki x + k p e, v = (ud − γ) (18)
|mmin | m
26 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

Note that v in (18) is essentially the inversion process for The first and second derivatives of the tracking error e are
the piecewise linear hysteresis model, and it requires tracking
ė = ẏ − ẏr = x2 − ẏr (26)
which linear segment the hysteresis characteristic is on at each
time instant. Such an assumption is standard in hysteresis
inversion as the past history of v is available. We further ë = −a1 x2 − a0 x1 + a0 (ud + d) − ÿr (27)
note that (18) requires knowing the slope and intercept of the
current linear segment; when such knowledge is not precise, From (26) and (27), we obtain
we can represent it as follows. Let us denote the corresponding ṡ = ë + σ1 ė + σ2 e
slope of the plant hysteresis Γ p by m p and the intercept by γ p
= −a1 x2 − a0 x1 + a0 (ud + d) − ÿr
with parameter uncertainties ∆m and ∆γ , where m p = m + ∆m
and γ p = γ + ∆γ , which implies +σ1 (x2 − ẏr ) + σ2 (x1 − yr ) (28)

u = (m + ∆m )v + (γ + ∆γ ) (19) Now, we may write the control ud as the sum of an


equivalent component ueq and switching component us control
By substituting ud and v from (18) into (19), we express u as
ud = ueq + us (29)
m + ∆m m∆γ − γ∆m where,
u= [gyr + ki x + k p (yr −Cz)] + (20)
m m us = −β sgn (s) (30)
The singularly perturbed closed-loop system, obtained by
inserting u from (20) into (16), is given by 1
ueq (t) = ([a1 − σ1 ]x2 + [a0 − σ2 ]x1 + ÿr
a0
+σ1 ẏr + σ2 yr ) (31)
ẋ =yr −Cz
k p (m + ∆m ) Using (29), we can write the bound on d as
m + ∆m
ε ż =[A − BC]z + B (g + k p )yr
m m |d(t)| ≤ k1 + k0 us (t) + ueq (t)
m + ∆m m∆γ − γ∆m
+B ki x + B( ) (21) ≤ k1 + k0 |us (t)| + k0 ueq (t) (32)
m m
Define
The solution e(t) of (21) is derived in [42] and expressed as
φ (|ueq (t)|) = k1 + k0 |ueq (t)| (33)
e = e0 + εeε + εωeεω + O(ε 2 ) (22) which implies
where e0 , eε , and eεω are bounded uniformly in ε and ω. In |d(t)| ≤ φ (|ueq (t)|) + k0 |us (t)| (34)
other words, |e0 | < k1 , |eε | < k2 , and |eεω | < k3 , where k1 , k2 ,
and k3 are some positive constants independent of ε and ω. Substituting (31) and (29) in (28), we obtain
The contribution from the eε term can be ignored because it ṡ = a0 (d(t) + us (t)) (35)
is multiplied by a small number ε. The term eεω becomes
significant at high frequency when its coefficient εω is not For the stability analysis, we define a Lyapnov function V =
1 2
small. Then e0 (t) becomes 2 s . Then by taking the derivative of V , we obtain

ω2 V̇ = sṡ
e0 (t) = A0i ϕ + B0i Ac sin(ωt) + A12 ψ (23)
A20 + ω 2 = a0 s (d(t) + us (t))
It is noted that for a sufficiently high frequency (|A0 |  ω  ≤ a0 (|s(t)| |d(t)| + s(t)us (t))
2
1/ε), A2ω+ω 2 becomes constant and almost independent on (36)
0
frequency. Now we substitute the bound on d from (34) into (36)
V̇ ≤ a0 (|s| [φ (|ueq (t)|) + k0 |us |] + sus )
B. SMC Controller
≤ a0 (|s| [φ (|ueq (t)|) + k0 β ] − β |s|)
The second-order linear system defined in (1) will be used
= a0 (− [1 − k0 ] β + φ (|ueq |)) |s|
to explain the idea of design the system. The sliding surface
is formed as follows (37)
Therefore V̇ < 0 for
s = ė + σ1 e + σ2 z (24)
φ (|ueq (t)|)
where z is the integration of the error, which is added to reduce β> (38)
1 − k0
the DC component of the error due to the shift in the hysteresis
Following the discussions in Section III, we assume 0 < k0 <
loop when the tracking error signal enters the boundary layer. φ (|ueq (t)|)
σ1 and σ2 are positive constants. 1. By choosing β = 1−k 0
+ β0 for some small positive
constant β0 , we guarantee that any trajectory starting away
ż = e (25) from the surface s = 0 reaches it in finite time. Note that β
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 27

The reference model is chosen to be


200
Gm = (45)
s + 200
In MRAC, the desired behavior of the system is described by a
linear system with a transfer function Gm (s) and is driven by a
reference input r(s). The control law C(s, θc ) is then developed
so that the closed-loop plant has a transfer function equal to
Gm (s). Fig. 9 shows the structure of Model-Reference-Control
(MRC). The plant transfer function is G(s, θ p ), where θ p is a
vector of all the coefficients of G(s).
The controller is designed such that
Fig. 9: The MRAC control system preceded by the inverse operator. yP ym (s)
= Gm (s) = (46)
r r(s)
depends on the equivalent control component ueq and hence This requires G(s) and Gm (s) to satisfy certain assumptions.
on the state and time variables. This transfer function matching of (46) guarantees that the
With the signum function in the switching controller (30), tracking error e1 = y p − ym converges to zero for any given
the error would asymptotically converge to zero. The signum input signal r. If the plant parameter θ p∗ is known, then the
function, however, results in chattering in practice. Therefore, controller parameter θc∗ can be calculated as a function of
we replace the signum function with the saturation function them and the controller C(s, θc ) is implemented. When these
given by parameters are unknown, they are replaced by their estimates
θ p and θc , leads to adaptive control scheme referred to as
us = −β sat (s/µ) (39) indirect or direct MRAC.
We can meet the control objective if we choose the con-
where µ determines the size of the boundary layer.
troller parameters such that the closed-loop poles are stable
To analyze the performance of the control with the switching
and the closed-loop transfer functions Gc (s) = Gm (s).
component (39), we examine the derivative of the Lyapunov
The design of Lyapunov function for a system with rela-
function V = 12 s2 outside the boundary layer.
tive degree n = 2 requires Gm (s) to be Strictly-positive-real
Plugging (39) into the first inequality (37), we have
SPR. In [43], the system is modified by using the identity
V̇ ≤ a0 (|s|[φ (|ueq (t)|) + k0 | − β sat(s/µ)|] (s + p0 )(s + p0 )− 1 for some p0  0. We can follow steps
−β ssat(s/µ)) (40) similar to Theorem 5.4.3 in [43] to show the boundedness
of the tracking error. The only difference that we added a
Then for |s| ≥ µ, we get bounded disturbance d ∈ L∞ , and we show that the tracking
V̇ ≤ a0 (− [1 − k0 ] β + φ (|ueq |)) |s| (41) error is bounded.
The MRAC scheme is summarized by the equations:
Thus, with the same β in (38) , |s| will decrease until it enters
the boundary layer |s| ≤ µ and never leaves it thereafter. ud = θ T w = θ1T w1 + θ2T w2 + θ3 y p + θ4 r (47)

C. MRAC Controller w˙1 = Fw1 + gud (48)


In the following part, we discuss the design of a Model
Reference Adaptive Control Scheme in the case of the
w˙2 = Fw2 + gy p (49)
presence of plant linear dynamics. The design is based on the
assumption that inversion of the hysteresis model is affecting where w1 0 = 0 and w2 0 = 0. The functions F and g are the
the parameter of the linear part and make them uncertain. The state realization of α(s)
Λ(s) .
design of MRAC control will take care of these uncertainties. Because the order of the plant is n = 2, we used α(s) = 1. In
addition, Λ(s) was chosen to be monic, Hurwitz with degree
The linear dynamics are expressed as n − 1 and it contains the zeros of the reference model as a
Z p (s) factor. Consequently, Λ(s) was chosen to be Λ(s) = s + 2.
G(s) = k p (42)
R p (s) The adaptive law, based on the gradient algorithm, is given
by
The transfer function of the reference model is expressed in
the same form as kp
θ̇ = −Γ.e.w.sgn( )
Zm (s) km
Gm = km (43) = −Γ.e.w (50)
Rm (s)
The linear dynamics are identified as where e = yp − ym and k p and km are the high frequency gains
a0 of the plant and reference model respectively. Γ is a diagonal
G(s) = 2 (44) matrix representing the adaptation gain. Based on Theorem
s + a1 s + a0
28 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

80 10
Operator Input Non feedforward method
feedforward (Scheme 1)
Position ( µ m )
60 Operator Output

40
5

Error ( µ m)
20
0
0
0 20 40 60 80 100 120

0 −5
Inversion Error ( µ m)

−1

−2
−10
−3 0.4 0.42 0.44 0.46 0.48 0.5
Time ( s )
−4

−5
Fig. 11: PI simulation results of the tracking error with and without forward-
0 20 40 60 80 100 120 path compensation.
Time ( s )

Fig. 10: The output and inversion error of a perturbed PI operator. 60

Position (µ m)
position
40 reference
5.4.1 of [43], the above MRAC scheme has the following 20
property; all signals in the closed-loop plant are bounded, and 0
the tracking error e1 converges to zero asymptotically with
time for any reference input r ε L∞ . 0 0.1 0.2 0.3 0.4
Error (µ m)

V. S IMULATION R ESULTS 0.5

The simulation is based on the parameters identified 0


experimentally for a commercial nanopositioner. The −0.5
hysteresis is modeled by PI operator with 5 play operators 0 0.1 0.2 0.3 0.4
with the threshold vector r = [0, 0.6364, 1.2727, 2.54, 4.45]T 4
x 10
and the vector of weights for the operator is 1
s(µ/s)

wT = [5.88, 1.58, 0.47, 0.98, 0.4]. The linear dynamics 0


are fitted with a second-order system with parameters
a1 = 1.7168 × 108 and a0 = 5743. −1
0 0.1 0.2 0.3 0.4
Simulation results are based on perturbing the model such Time ( s )
that they result in an error close to the one comes from
cascading the inverse-operator with the actual positioner Fig. 12: The SMC simulation results on tracking non-sinusoidal reference:
hysteresis. Each component on the weight vector is perturbed position, tracking error and s trajectories.
by a maximum value of ∆w,max = 0.15. Fig. 10 shows the
error results from cascading the inverse-operator with the
1
perturbed hysteresis operator. γ = 0.5 for all parameters. The filters are s+200 . The model
For PI controller, Fig. 11 depicts the tracking errors reference is a first-order linear plant. This simplify the
for a reference of 50 Hz for the general feedback scheme tracking as the bandwidth of the plant is large and can
(without the feedforward term) and for the method when the approximated by a first order system. The reference model is
b0
feedforward term is added, where perfect hysteresis inversion selected as Gm = s+b 0
, where b0 = 104 .
is assumed. From these results we confirm that feedforward- Fig. 14 shows the results of tracking a triangular reference
augmented feedback outperforms feedback alone. of 25 µm traveling range and 5 Hz. The tracking error will
To validate the proposed SMC method, non-sinusoidal be evaluated in terms of the maximum value of the absolute
reference signals are generated using van der Pol oscillator error ( Max |e(t)|) at steady state. The system has a good
with traveling ranges of 20 µm to 50 µm. In our design we tracking with ( Max |e(t)| = 0.8µ m).
have k0 = 0.1282 and k1 = 10.53 Fig. 12 depicts the output, In Fig. 15, a comparison between the tracking of the system
tracking error and s for a reference that has a “fundamental when the inverse-operator is inserted and removed from the
frequency” of 15 Hz. From these results we notice that the closed-loop system. It shows an improvement when the input
trajectories enter the boundary layer in a very short time is a pure sin-wave, where the tracking error is improved by
and the output tracks the reference after a few cycles. The decreasing from 2.5µm to 0.6µm.
maximum tracking error for this signal is 1.47%. Fig. 13
shows that the tacking error is increased as the frequency Finally, we made a comparison for three control methods
increases. in Table I. The maximum value of the of the tracking error
For MRAC control, The adaptation constants were chosen (Max |e(t)|) is compared at 10–400 Hz frequencies. A refer-
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 29

50 Hz a) without inverse−operator

Tracking Error ( µ m)
5
0.5
Error ( µ m )

0
0

−0.5

−5
−1 0.1 0.15 0.2 0.25 0.3
0.3 0.35 0.4 0.45
b) with inverse−operator

Tracking Error ( µ m)
15 Hz 1

0.2
0.1
0
Error ( µ m )

0
−0.1
−0.2 −1
0.1 0.15 0.2 0.25 0.3
−0.3 Time ( s )
0.3 0.35 0.4 0.45 Fig. 15: MRAC simulation results: tracking errors for a sine- reference with
Time ( s ) and without inserting the inverse-operator.

Fig. 13: SMC simulations: Tracking errors for references with different
“fundamental frequencies”. has almost constant tracking error in the frequency band from
80–200 Hz. Yet, we cannot say SMC is the best method in
general because there are some other restrictions. For instance
30 SMC has chattering phenomena when it is run over 100 HZ
Reference in real setup. Moreover, high order derivative of the input is
Position

20 Output required in the design of SMC which can be obtained when


it is sinusoidal, but is hard to obtain that with arbitrary shape
10 signals or even with simple triangular waveforms. This is not
required with MRAC control.
0
0 0.2 0.4 0.6 0.8 1
TABLE I: S IMULATION R ESULTS P ROVIDING A C OMPARISON B ETWEEN
M AXIMUM T RACKING E RRORS FOR PI, H∞ , AND N EURAL SMC C ONTROL
Tracking Error

20
freq. PI MRAC SMC
Max |e(t)| Max |e(t)| Max |e(t)|
10 10 Hz 0.03 0.2 0.05
20 Hz 0.065 0.32 0.07
0 50 Hz 0.15 0.45 0.08
80 Hz 0.22 0.5 0.09
100 Hz 0.265 0.58 0.09
−10 200 Hz 0.4 0.8 0.085
0 0.2 0.4 0.6 0.8 1 400 Hz 0.7 0.9 0.08
Time ( s )

Fig. 14: MRAC simulation results: position, and tracking error.


VI. C ONCLUSION
In this paper, Feedforward and feedback control methods are
ence sinusoidal signal of amplitude 50 µm is applied to the compared when they are used for controlling smart materials.
system. We checked how this error scales with frequency. We These materials exhibit some nonlinear phenomenon such
notice that below 10 Hz the PI-controller outperforms the other as hysteresis. A PI operator, which is used to model the
controllers. This is expected because the PI-controller has a hysteresis and its inverse is implemented in the forward path
very high gain at frequencies close to zero, which dominates to reduce the disturbance to the linear plant. The feedforward
uncertainties. The MRAC has the larger tracking error at all method reduces the tracking error appreciably, but integrating
frequencies presented in Table I. It is also slower in response it with feedback methods improves the results as required in
as we see in Fig. 14. On the other hand, the Sliding-Mode many nano-positioning applications.
controller (SMC) outperforms the other two methods when the Each feedback method has its advantages and drawbacks.
frequency is more than 10 Hz. It is also notable that SMC also The conventional Proportional-Integral PI control is the
30 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

simplest and gives good results at low frequency. For higher [20] S. Bashash and N. Jalili, “Robust multiple frequency trajectory tracking
frequency, we discussed the robust SMC control which control of piezoelectrically driven micro/nanopositioning systems,” IEEE
Transactions on Control Systems Technology, vol. 15, no. 5, pp. 867–
outperforms the PI controller. Yet, it is hard to design SMC, 878, sept. 2007.
particularly, for high order system. The MRAC Adaptive [21] H. Liaw, B.Shirinzadeh, and J. Smith, “Enhanced sliding mode motion
method is also examined and shows moderate performance tracking control of piezoelectric actuators,” Sensors and Actuators A,
vol. 138, pp. 194–202, 2007.
results. Adding the inverse-operator is not improving the [22] C. Jan and C. Hwang, “A nonlinear observer-based sliding-mode control
results for it as in the robust methods. for piezoelectric actuator systems: Theory and experiments,” Journal of
Although, the SMC has the best tracking results, but there the Chinese Institute of Engineers, vol. 27, no. 1, pp. 9–22, 2004.
[23] X. Tan and H. K. Khalil, “Two-time-scale averaging of systems involving
are other things which is should be considered. For instance, operators and its application to adaptive control of hysteretic systems,”
we cannot operate it with any arbitrary-shape waveform. It in Proceedings of the 2009 American Control Conference, 2009, pp.
also produces chattering when it is experimentally run at 4476–4481.
[24] K. Kuhnen and Janocha, “Adaptive inverse control of piezoelectric
high frequency. As an extension to this work, I suggest for actuators with hysteresis operator,” in Proceedings of the European
future to compare these methods under different experimental Control Conference, Karsruhe, Germany, 1999, paper F 0291.
setup. This makes all the miscellaneous factors available for [25] J. Zhong and B. Yao, “Adaptive robust precision motion control of a
piezoelectric positioning stage,” IEEE Transactions on Control Systems
the designer to determine their best choice, depending on the Technology, vol. 16, no. 5, pp. 1039 –1046, 2008.
application and the working environment. [26] Y. Wu and Q. Zou, “Iterative control approach to compensate for
both the hysteresis and the dynamics effects of piezo actuators,” IEEE
Transactions on Control Systems Technology, vol. 15, no. 5, pp. 936–
R EFERENCES 944, 2007.
[27] J. Muir, “On the recovery of iron from overstrain,” Philosophical
[1] C.J and C. P, “Piezoelectric and allied phenomena in rochelle salt,” Transactions of the Royal Society of London Series A, vol. 193, pp.
Comput Rend Acad Sci, vol. 91, pp. 294–297, 1880. 1–46, 1900.
[2] X. Tan and J. S. Baras, “Modeling and control of hysteresis in magne- [28] F. G. Baily, “The hysteresis of iron and steel in a rotating magnetic field,”
tostrictive actuators,” Automatica, vol. 40, no. 9, pp. 1469–1480, 2004. Philosophical Transactions of the Royal Society of London Series A, vol.
[3] R. Smith, Smart Material Systems: Model Developments. Society of 187, pp. 715–746, 1896.
Industrial and Applied Mathematics, 2005. [29] M. Krasnoselskii and A. Pokrovskii, “Periodic osccilatoin in systems
[4] S. Devasia, E. Eleftheriou, and S.O.Moheimani, “A survey of control with relay nonlinearities,” Akademiia Nauk SSSr, Doklady, vol. 216, pp.
issues in nanopositioning,” IEEE Transactions on Control Systems 733–736, 1974.
Technology, vol. 15, no. 5, pp. 802 –823, sept. 2007. [30] ——, Systems with Hysteresis. New York: Springer-Verlag„ 1989.
[5] D. Croft, G. Shedd, and S. Devasia, “Creep, hysteresis, and vibration [31] R. Mrad and H. Hu, “A model for voltage-to-displacement dynam-
compensation for piezoactuators: atomic force microscopy application,” ics in piezoceramic actuators subject to dynamic-voltage excitations,”
in Proceedings of the 2000 American Control Conference, 2000, pp. IEEE/ASME Transactions on Mechatronics, vol. 7, no. 4, pp. 479–489,
2123–2128. 2002.
[6] K. Kuhnen and Janocha, “Inverse feedforward controller for complex [32] X. Tan and J. Baras, “Adaptive identification and control of hysteresis
hysteretic nonlinearities in smart-material systems,” Control and Intel- in smart materials,” IEEE Transactions on Automatic Control, vol. 50,
ligent systems, vol. 29, no. 3, pp. 74–83, 2001. no. 6, pp. 827–839, 2005.
[7] G. Tao and P. Kokotovic, “Adaptive control of plants with unknown [33] R. Iyer and X. Tan, “Approximate inversion of the preisach hysteresis
hystereses,” IEEE Transactions on Automatic Control, vol. 40, no. 2, operator with application to control of smart actuators,” IEEE Transac-
pp. 200–212, 1995. tions on Automatic Control, vol. 50, no. 6, pp. 798–810, 2005.
[8] Q. Xu and Y. Li, “Radial basis function neural network control of an XY [34] H. Hu and R.Mrad, “On the classical preisach model for hysteresis in
micropositioning stage without exact dynamic model,” in Proceedings of piezoceramic actuators,” Mechatronics, vol. 13, no. 2, pp. 85–94, 2002.
the 2009 IEEE/ASME International Conference on Advanced Intelligent [35] H. Hu, H. Georgiou, and R. Ben-Mrad, “Enhancement of tracking ability
Mechatronics, july 2009, pp. 498–503. in piezoceramic actuators subject to dynamic excitation conditions,”
[9] D. Youla and J. Bongiorno, “A feedback theory of two-degree-of- IEEE/ASME Transactions on Mechatron, vol. 10, no. 2, pp. 230–239,
freedom optimal wiener-hopf design,” IEEE Transactions on Automatic 2005.
Control, vol. 30, no. 7, pp. 652–665, Jul 1985. [36] D. Carnevale, S. Nicosia, and L. Zaccarian, “Generalized constructive
[10] I. Mayergoyz, “Mathematical models of hysteresis,” IEEE Transactions model of hysteresis,” IEEE Transactions on Magnetics, vol. 42, no. 12,
on Magnetics, vol. 22, no. 5, pp. 603–608, 1986. pp. 3809–3817, 2006.
[11] P. Ge and M. Jouaneh, “Generalized preisach model for hysteresis [37] R. Mrad and H. Hu, “Dynamic modeling of hysteresis in piezoceram-
nonlinearity of piezoceramic actuators,” Precision Engineering, vol. 20, ics,” in IEEE/ASME International Conference on Advanced Intelligent
no. 2, pp. 99–111, 1997. Mechatronics, 2001, pp. 510–515.
[12] M. Brokate and J. Sprekels, Hysteresis and phase transitions. Springer [38] J. W. Macki, P. Nistri, and P. Zecca, “Mathematical models for hystere-
Verlag, 1996. sis,” SIAM Review, vol. 35, no. 1, pp. 94–123, 1993.
[13] K. Kuhnen, “Modeling, identification and compensation of complex hys- [39] R. Iyer and M. Shirley, “Mathematical models of hysteresis,” IEEE
teretic nonlinearities - a modified Prandtl-Ishlinskii approach,” European Transactions on Magnetics, vol. 40, no. 5, pp. 3227–3239, 1986.
Journal of Control, vol. 9, no. 4, pp. 407–418, 2003. [40] M. A. Janaideh, S. Rakheja, and C.-Y. Su, “An analytical general-
[14] P. Ge and M. Jouaneh, “Tracking control of a piezoceramic actuator,” ized prandtl–ishlinskii model inversion for hysteresis compensation in
IEEE Transactions on Control Systems Technology, vol. 4, no. 3, pp. micropositioning control,” IEEE/ASME Trans. on Mechatron., vol. 16,
209 –216, 1996. no. 4, pp. 734–744, 2011.
[15] S. Valadkhan, K. Morris, and A. Khajepour, “Robust PI control of [41] R. Iyer and X. Tan, “Control of hysteretic systems through inverse
hysteretic systems,” in Proceedings of the 47th IEEE Conference on compensation: Inversion algorithms, adaptation, and embedded imple-
Decision and Control, 2008, pp. 3787–3792. mentation,” IEEE Control Systems Magazine, vol. 29, no. 1, pp. 83–99,
[16] M. Tsai, “Robust tracking control of a piezoactuator using a new ap- 2009.
proximate hysteresis model,” Journal of dynamic systems, measurement, [42] M. Edardar, X. Tan, and H. K. Khalil, “Closed-loop analysis for systems
and control, vol. 125, pp. 96–102, 2003. with fast linear dynamics preceded by hysteresis,” in Proceedings of the
[17] G. Schitter, P. Menold, H. Knapp, F. Allgöwer, and A. Stemmer, “High 2013 American Control Conference, to appear 2013.
performance feedback for fast scanning atomic force microscopes,” [43] P. Ioannou and B. Fidan, Adaptive Control Tutorial. SIAM, 2006.
Review of Scientific Instruments, vol. 72, no. 8, pp. 3320–3327, 2001.
[18] J. Slotine and W.Li, Applied Nonlinear control. Printce-Hall, 1991.
[19] J. Shen, W. Jywea, H. Chiang, and Y. Shub, “Precision tracking control
of a piezoelectric-actuated system,” Precision Engineering, vol. 32, pp.
71–78, 2008.
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 31

An Overview of Mobile Indoor Coverage: A Case


Study of a Local Mobile Operator
Maoloud Dabab1

Abstract—With the significant increase in large building de- indoor coverage is the main challenge for Almadar Aljadid
signs and rapid growth in indoor mobile traffic, indoor coverage and other mobile operators around the world, particularly due
solutions have become the highest priority for many operators. to the increasing size of buildings, as significant developments
As a result, wireless vendors and research centers have raced
to innovate new practical solutions for indoor coverage. This in new mobile technologies and mobile devices.
article discusses the importance of providing indoor coverage
for mobile operators, by focusing on Almadar Aljadid Company, The objectives of this article are: (1) to highlight the
a leading mobile operator in Libya. Moreover, it summarizes main reasons mobile operators care about improving indoor
the factors that lead to poor coverage inside buildings and coverage; (2) to explain the two fundamental concepts of
provides examples to clarify this issue in non-technical terms.
Additionally, the article describes several solutions that Almadar cellular system, which are network capacity and network
Aljadid has implemented, and presents Wi-Fi calling technology coverage; (3) to analyze the common causes of poor indoor
as a potential solution that will enable the Company to enhance coverage and summarize the impact of smart phones,
indoor coverage. frequencies and buildings; (4) to describe solutions for
Index Terms—Indoor Coverage, Mobile Indoor, DAS, Wi-Fi enhancing indoor coverage used by Almadar Aljadid
Calling, Indoor solution, Poor Coverage, Almadar Aljadid. Company; and (5) to present a simple comparison between
those solutions in terms of certain factors that operators
consider very important, including type of building, capacity
I. I NTRODUCTION
of the solution, difficulty of installation and cost.
OBILE operator companies must overcome numerous
M obstacles in order to provide satisfactory services.
According to several studies [1], [2], around 70-80% of all II. W HY I NDOOR C OVERAGE IS N ECESSARY
mobile traffic is located inside buildings. As a result, the
most important issues that mobile operators need to focus Mobile operators have varying perspectives on the necessity
on relate to the lack of indoor coverage – inside offices, of providing indoor coverage. Their points of views can be
high buildings, and floors located underground. Although it divided into two categories:
is difficult to provide exhaustive indoor mobile coverage,
providing subscribers with the ability to make high quality A. Customer Service
calls and to have reliable data connection everywhere inside Customer service is considered one of the most important
buildings can be regarded as complete coverage by the areas in which many telecom companies compete to provide
subscribers [3]. Indeed, providing a satisfied indoor coverage good services that satisfy their customers. As such, it is
is an important factor for attracting and retaining subscribers. considered key to these companies’ success. The leading
companies have studied the requirements of the market and
Providing indoor coverage from outdoor sites can be in have taken into account subscriber demands. Nowadays, the
some cases, but the outdoor sites must be installed near the most common customer complaint is poor indoor coverage.
buildings with proper transmitting power in order to provide Almadar Aljadid receives hundreds of complaints monthly
indoor services [4]. Overpower, however, can lead to network about low coverage and difficulty of connection2 some of
interference [2]. The necessity of providing multi-layer which are related to indoor environments. In response, Al-
cellular architectures is one of the strategies that is used to madar Aljadid is focusing on providing indoor coverage to
enhance mobile network coverage and full mobility for the satisfy customers.
traffic. In general network structures, there are three layers,
umbrella, macro and micro [5]. From this initial concept, the
indoor solutions have become increasingly popular. B. Market Forecast
According to Almadar Aljadid Company, a leading mobile The number of mobile subscribers has boomed in the
operator in Libya, the most common customer complaint is last 10 years [9], so there is a positive relationship between
poor coverage, specifically inside buildings . Thus, improving the number of subscribers and the amount of traffic. An
estimated 70-80% of the total traffic occurs inside buildings in
1 Maoloud Y. Dabab is Radio Network Planning and Optimization
developing countries [1], [2]. For this reason, mobile operators
Engineer – Almadar Aljadid Company, E-mail: m.dabab@almadar.ly;
dabab@pdx.edu. are competing to provide the best indoor environment services.
2 Resource: Almadar Aljadid Data Base with permission. Fig. 1 projects the future growth of traffic in global mobile
32 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

data traffic by region [10]. Fig. 2 illustrates the particular the system at the same time, the period of each connection, and
example of Almadar Aljadid’s western region network traffic the type of service that subscribers use. However, the traffic
during the period in the past. Due to the unstable situation in and how many subscribers the network can handle can be
Libya nowadays, it cannot be relied upon Almadar Aljadid estimated using several techniques. One technique that is used
Network to take live data because of the lack of the network to calculate Erlang (E) per subscriber in a GSM network is
stability. as follows [6]:
n×T
A= (1)
3600

A = Offered traffic from one or more users.


n = Number of calls per hour.
T = Average call in seconds.
E = Erlang = unit of measurement of traffic.

Using these techniques and the marketing target for


subscribers, planners who design mobile networks can
determine the network capacity. Then, the planners can define
Fig. 1: The growth of traffic in global mobile data by regions from 2014 to
2019. the number of base stations (BS) and their capacity. Each
BS has a certain number of transceivers (TRX). A TRX
can handle between 75-300 subscribers [7]. However, this
calculation is not necessary for existing networks in which
the traffic can be calculated more easily and accurately.
For instance, for Almadar Aljadid the average amount
of traffic per subscriber is about 33mE 1 . However, the
network planners still face the problem of estimating the
number of subscribers to calculate the target network capacity.
Therefore, they always add a margin ratio to avoid congestion.

B. Network coverage
The area in which cellular services are offered for sub-
scribers is known as network coverage. Telecom companies’
top priority is to provide good services to their subscribers, and
the most important service is providing coverage in all places.
Due to the increasing demand for coverage inside buildings,
companies tend to build in various indoor coverage solutions.
Fig. 2: Almadar Aljadid western region network traffic in quarters from 2009 In fact, while planners are designing the coverage network,
to 20101 .
they calculate the required signal strength in places that should
be covered, taking into account many factors. The required
signal strength (SSreq ) is calculated as shown below [8]:
III. T HE I MPORTANT C ONCEPTS OF C ELLULAR S YSTEMS
To understand the ecosystem of mobile indoor coverage, one SSreq = M Ssen + RFmarg + IFmarg + BL (2)
must understand two key concepts: the capacity of the mobile
network, and the level and quality of the signal strength. These
factors are relevant when seeking a suitable indoor solution. M Ssen = Mobile Station Sensitivity.
RF marg = Rayleigh Margin.
A. Network capacity IF marg = Interference Margin.
The general meaning of network capacity is the number BL = Body Loss.
of subscribers and the amount of traffic that can be handled
by the network at the same time. It is challenging to calculate
mobile traffic exactly because it depends on many factors such A big challenge for the planner is to provide contented
as the estimated number of subscribers who are connected to key performance indicators (KPIs), especially inside buildings.
Indoor solutions build on different factors, including amount
1 Resource: Almadar Aljadid Data Base with permission. of traffic inside crowded buildings like airports and stadiums.
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 33

Another factor is the strength of the indoor signal, where the B. Frequencies
outdoor sites do not cover the inside of large buildings like In addition to its impact on path loss, frequency has a direct
hospitals and hotels effectively [8]. influence on coverage range. To understand why frequency
leads to poor coverage, the article discusses the concept of
wavelength, which is the distance between peaks of a radio
IV. R EASONS FOR P OOR C OVERAGE I NSIDE B UILDINGS
wave, and it is measured in meters. There is an inverse rela-
To find a suitable solution for poor indoor coverage, the tionship between frequency and wavelength, which determines
network planners must understand what weakens coverage. by the following equation [2]:
Among other factors, signal obstacles stand out as the main Speed
cause. In this section, these factors can be separated into two W avelength(λ) = (4)
F requency
categories: the obstacle to the signal, which is buildings, and
characteristics of the signal, which are frequencies and smart Whereas speed of light = 3 × 108 m/s (in space).
phones.
Therefore, the wavelength of f = 900M Hz is:

A. Buildings
3 × 108 m/s
Urban development has a significant impact on the quality W avelength = = 0.33m (5)
900M Hz
of coverage, in particular buildings’ structures. Heavy concrete On the other hand, the wavelength of f = 1800M Hz is:
buildings that lead to a high ratio of reflection disrupt the
signal strength of mobile phones. Indoor path loss is clearly 3 × 108 m/s
affected by several factors and is a function of time and W avelength = = 0.167m (6)
1800M Hz
location. The value of International Telecommunication Union The distance of the Global System of Mobile GSM
(ITU) path loss (L) for an indoor propagation model can be 900 coverage is larger than that of GSM 1800. From this
estimated with this equation [8]: example, it is obvious that when the frequency increases,
the wavelength decreases. Therefore, the range of coverage
L = 20LOG(f ) + N Log(d) + Lf (n) − 28dB (3) depends on frequency of the technology. The first frequency
band Almadar Aljadid Company used was 900MHz. They
have since increased the frequency to 1800MHz and may
increase it in the near future to 2.1GHz. The frequency has
N = The distance power loss coefficient. been gradually climbing, resulting in a lower coverage range.
f = Frequency in MHz. Additionally, it is obvious from Equation (3) that the path
loss rises slightly as frequency increases.
d = Distance in meters.
Lf (n) = The floor penetration loss factor.
n = Number of floors between the RX and TX. C. Smart phones
It is obvious from Fig. 3 that smart phone usage is climbing
rapidly [11], and planners must take this into account in their
The penetration loss factor (PLF) depends on the type of wall calculations. The mobile station (MS) has a significant role
and its thickness. Therefore, buildings with solid walls and in determining signal strength, in which each type of MS has
a greater number of floors may cause increased path loss. different varying output power and sensitivity [8].
Moreover, PLF is a function of frequency, high and t of the
building materials. Table I below shows specific examples The new generations of phones have a small patch
of PLF influence [10]. It clear that PLF is higher at thick antenna that has a negative affect on their mobile coverage
walls, which are made of reinforced concrete, than thin board performance. Additionally, the greater number of applications
dividing walls with all frequencies. that are open, the more coverage is reduced [12]. Table II
below shows a simple comparison between two groups of
TABLE I: P ENETRATION L OSS FACTOR FOR T WO D IFFERENT T YPES OF
WALL AS A F UNCTION OF F REQUENCY users in terms of the problems they face. The first group
is smart phone users and the second is other cell phone
Frequency Loss for thin walls Loss for thick walls users [13].
(GHz) (dB) (dB)

2 3.3 10.91
V. S OLUTIONS TO I MPROVE I NDOOR M OBILE C OVERAGE
3.5 3.4 11.4
It follows from the previous discussion that buildings have
5 3.4 11.8 been the most complex problem, and most variable factor
affecting indoor coverage. This is because all buildings are
different. Therefore, the solution will be unique for each
34 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

important to cover, some Almadar Aljadid sale centers, and


various rural farms are examples of locations covered by this
solution1 .

Fig. 3: The growth of smart phone shipment forecast worldwide from 2006
to 2020 with estimation of Compound Annual Growth Rate as 16% in the
future. Fig. 4: A simple diagram shows the main parts of an indoor RF repeater and
an outdoor RF repeater.
TABLE II: S MART P HONE OWNERS C ONFRONT C HALLENGING M OBILE
P ROBLEMS % IN EACH G ROUP WHO H AVE E NCOUNTERED M OBILE P HONE
P ROBLEMS AT L EAST W EEKLY
B. Small cells
Smartphone Other cell Micro cells, pico cells and femtocells are types of small
owners owners
cells, which have low output power and a limited range of
Dropped calls (among cell owners) 35%* 28% coverage, installed to cover small areas, especially the spots
Unwanted sales, marketing calls that cannot converged by macro cells to provide good indoor
(among cell owners) 26% 23% coverage, such as enclosed malls and playgrounds. However,
Spam, unwanted text (among own- the range of the small cells are limited in comparison to
ers) 29%* 20%
the macro cells, and they have a low capacity. Therefore,
Slow download speeds (among mo-
bile internet / email / app users) 49%* 31% they are used to enhance indoor coverage for areas with
* Denotes statistic a GYU lly significant difference limited space and low traffic capacity. Fig. 5 illustrates the
differences between macro cell and micro cells, pico cells and
femtocells [15]. Almadar Aljadid once had many locations
covered by micro cells, but they have decreased, having been
case, regardless of common factors, such as sensitivity of
replaced in many locations with macro cells. However, there
phones and frequencies that are used in networks. The indoor
are still places that have small cells, such as Zwara Airport
solutions, which have been used by Almadar Aljadid and will
and certain oil fields1 .
be used in the future, are:

A. Repeater
A Radio Frequency (RF) repeater can provide indoor cov-
erage for small buildings or enclosed areas. The RF repeater,
which consists of a receiver antenna, cable, signal processor
and transmitter antenna, amplifies the signal from the Base
Station (BS) to the Mobile Station (MS), and uses the capacity
of the BS. There are two types of RF repeaters, indoor and
outdoor, and the main difference between them is the location
of the transmitter antenna. In addition, there are many repeater
brands with different characteristics and coverage range, and
some of them have higher performance than others [14]. Fig. 5: Different kind of small cells compared with a macro cell.
Fig. 4 diagrams the basic indoor RF repeater and outdoor
RF repeater. This solution has been used in various locations
throughout Libya. For instance, Almadar Aljadid installed C. Distribution Antenna System
indoor and outdoor RF repeaters in the house of high-level
executives, small government buildings that cannot be cov- The most effective solution for in-building coverage to
ered by outdoor cells, and in private buildings that are very crowded buildings is a Distribution Antenna System (DAS),
which distributes RF through the building to supply all areas
1 Resource: Almadar Aljadid Data Base with permission. with good signal strength. The DAS has a network of cables,
ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016 35

couplers, splitters and antennas as seen in Fig. 6. DAS can be VI. C OMPARISON B ETWEEN S OLUTIONS
passive, using passive components without any amplification Operators consider technical solutions from several angles,
of signal, or it can be active, using optical distribution with unlike subscribers who look only at the quality of service.
a number of remote units, or hybrid, which is a combination Table III displays a simple comparison between the indoor
of passive and active, and may be used in some solutions. solutions in terms of building types, the capacity of the
The advantage of DAS is its ability to support many operators solution, the difficulty of installation, and the cost, which are
in the same indoor network, as well as provide satisfactory the most important for operators.
services to customers with high capacity traffic [16], [17].
Almadar Aljadid has constructed many DAS sites that cover TABLE III: C OMPARISON B ETWEEN THE I NDOOR S OLUTIONS IN T ERMS
OF S IGNIFICANT FACTORS TO O PERATORS
large buildings such as the Corinthia Hotel, the Tripoli Tower,
and the Tibesti Hotel. It is currently seeking to extend its

Installation
capacity
building
Type of

Notes
indoor network and include other buildings requiring indoor

Cost
coverage in the near future1 .
Some types have

Repeater
bad effect on
Small N/A Easy Low network
performance

Small
Short range of

cell
Open plan Low Easy Medium
structure coverage

DAC
Very common in
Large High Hard High
large buildings

Calling
With

Wi-Fi
Internet High N/A N/A New technology
connection

Fig. 6: A simple diagram of distribution antenna system network inside


VII. C ONCLUSION
buildings from different angel. The need for indoor coverage emerges mainly inside high
buildings, large malls and basements where subscribers suffer
from poor coverage. This article discussed the concept of
D. Wi-Fi calling mobile indoor coverage in terms of challenges and solutions
as well as the importance of providing it in the first place.
Currently, Wi-Fi Internet connections are in use everywhere,
Like other mobile operators, Almadar Aljadid Company has
especially inside buildings. Mobile operators can take
been using a variety of methods to solve this problem, and
advantage of this feature to enhance their mobile services.
it continues researching and exploring new technologies.
The most recent solution to improving indoor environment
This an overview also serves as prelude to a series of more
is Wi-Fi calling technology. This simple solution involves
specialized and advanced future studies on indoor coverage
making phone calls through Wi-Fi connections using voice
and related topics.
over IP when the mobile network is not available. However,
the main issue with this technology is that the quality of calls
depends completely on the condition of internet connection
R EFERENCES
services. In addition, this solution is only applicable to phones
that support Wi-Fi calling technology, which is problematic [1] J. S. Seybold, “Indoor Propagation Modeling,” Introduction to RF
propagation, Hoboken, N.J.: Wiley, pp. 208–211, 2005.
for many subscribers who still use older generations of [2] M. Tolstrup, "Indoor Radio Planning" Indoor radio planning a practical
phones [18]. For some mobile operator companies that use IP guide for GSM, DCS, UMTS, HSPA and LTE, 2nd ed, Chichester, United
networks as backhaul, Wi-Fi calling is a very good solution. Kingdom: John Wiley & Sons, ch. 3, pp. 27, 2011.
[3] M. Tolstrup, Indoor Radio Planning A Practical Guide for 2G, 3G and
For instance, T-Mobile, an operator in the U.S., offers a 4G, 3rd ed. Hoboken: Wiley, 2015.
Wi-Fi calling service, and it provides voice calls and short [4] Y. Qian, Z. Guo, R. Fan, H. Wang, J. Liu, Y. Yan, X. Shen, and Z. Hu,
message services to its subscribers. T-Mobile hosts around 40 “Improving outdoor to indoor coverage by use of TD-LTE in-band
relay,” 2013 IEEE 24th Annual International Symposium on Personal,
million Wi-Fi calls per month [19]. Indeed, Almadar Aljadid Indoor, and Mobile Radio Communications (PIMRC), pp. 2658–2662,
network is not currently ready to provide this technology 2013.
because it requires voice over IP. Also, there are no high [5] E. Kadir and M. Baba, “Strategy to improve the indoor coverage for
mobile station,” Student Conference on Research and Development, pp.
speed internet connections in Libya. These limited resources 410–413, 2002.
prevent Almadar Aljadid from taking advantage of the Wi-Fi [6] GSM System Survey, student text, 123 33 21 R3B, Ericsson Co., pp.
calling for indoor coverage, but as resources expand, the 210-212, 1998.
[7] C. Seltzer, “Indoor Coverage Requirements and Solutions,” Antennas
company may be able to utilize this technology. and Propagation for Future Mobile Communications, No. 5927503, pp.
3/1–3/4, February 1998.
[8] GSM plannng workshop, student text, 123 33 15 R3B, Ericsson Co., pp.
1 Resource: Almadar Aljadid Data Base with permission. 231-233, 1999.
36 ALMADAR JOURNAL FOR COMMUNICATIONS, INFORMATION TECHNOLOGY, AND APPLICATIONS, VOL. 03, NO. 01, APRIL 2016

[9] Y. Yamao, H. Suda, N. Umeda, and N. Nakajima, “Radio access network


design concept for the fourth generation mobile communication system,”
VTC2000-Spring. 2000 IEEE 51st Vehicular Technology Conference
Proceedings (Cat. No.00CH37026), vol. 3, pp. 2285–2289, 2000.
[10] H. Nafea, F. Zaki, H. Moustafa, "Penetration Loss of Walls and Data
Rate of IEEE802.16m WiMAX including Adaptive Modulation and
Coding (AMC)," Science and Education Publishing, Vol. 1, Issue. 1,
pp. 29-36, 2013.
[11] T. Danova, "The Global Smartphone Report: These are the
companies and regions that will drive the next billion smartphone
shipments,” Business Insider, [Online]. Available:http://www.business
insider.com/global-smartphone-market-forecast-vendor-platform-
growth-2015-6, 19 June, 2015.
[12] "Top 10 reasons your smartphone signal fails," Arabian Business, [On-
line]. Available:http://www.arabianbusiness.com /photos/top-10-reasons-
your-smartphone-signal-fails-568750.html, 23 October, 2014.
[13] J. Boyles, L. Rainie, "Mobile Phone Problems," Pew Research Center
Internet Science Tech RSS, [Online]. Available:http://www.pewinternet
.org/2012/08/02/mobile-phone-problems/., 2 August, 2012.
[14] R. L. Durrant, L. Scott , and J. K. Reece, “RF signal repeater in mobile
communications systems,” Patent US6839539 B2, 4 January, 2005.
[15] “HetNet Synchronization”, Synchronizing HetNets for 4G LTE and
beyond • Qulsar, [Online]. Available: http://qulsar.com/Applications
/Telecom_and_Networks/HetNets.html.
[16] H. Beijner, “The importance of in-building solution in third-generation
networks,” Ericsson Co., Review No.2, pp. 90-97, 2004.
[17] Nokia Co., Indoor Deployment Strategies White Paper, Finland: Nokia
Solutions and Networks Oy, C401-00991-WP-201405-1-EN, June 2014.
[18] S. Hasan, N. Siddique, S. Chakraborty, “Femtocell versus WiFi - A
survey and comparison of architecture and performance,” 1st Interna-
tional Conference on Wireless Communication, Vehicular Technology,
Information Theory and Aerospace & Electronic Systems Technology.
pp. 916-920, May 2009.
[19] "T-Mobile Android Smartphones to Get WiFi Calling 338304." eWeek
6 October. 2010. Academic OneFile. Web, 22 November, 2015.

Potrebbero piacerti anche