INTERNATIONAL JOURNAL OF INTELLIGENT SYSTEMS AND APPLICATIONS (IJISA) ISSN Print: 2074904X, ISSN Online: 20749058 EditorinChief Prof. Kohei Arai, Saga University, Japan
Associate Editors Prof. Prabhat Kumar Mahanti, University of New Brunswick, Canada Prof. Mohammed AboZahhad AboZeid, Assiut University, Egypt Prof. Ayman A. Aly, Taif University, Saudi Arabia Prof. Santosh Kumar Nanda, Eastern Academy of Science & Technology, India Prof. M.S. Khireddine, Batna University, Algeria
Members of Editorial and Reviewer Board Prof. Kusum Deep Indian Institute of Technology Roorkee, India Prof. Anupam Agrawal Indian Institute of Information Technology, India Dr. T. Kishore Kumar National Institute of Technology Warangal, India Dr. Van Dinh Nguyen Ha Noi University of Agriculture, Vietnam Prof. Amar Partap Singh Sant Longowal Institute of Engineering and Technology, India Dr. Mohamed Zellagui University of Batna, Algeria Prof. D.V.Pushpa Latha GokarajuRangaraju Institute of Engineering and Technology, India Prof. Nikolay N. Karabutov Moscow state engineering Russian
university,
Prof. B.K. Tripathy VIT University, India Dr. Farzin Piltan Institute of Advance Science TechnologyIRAN SSP, Iran
Dr. Imtiaz Hussain Khan King Abdulaziz University, Saudi Arabia
and
Dr. Hossam Eldin Mostafa Attia Jubail Industrial College (JIC), Saudi Arabia Dr. Sudip Kumar Sahana Birla Institute of Technology, India Dr. Avinash J. Agrawal Shri Ramdeobaba College of Engineering and Management, India Dr. Hussein Jaddu Al Quds University, Palestine
Dr. Chinmaya Kar Honeywell Technology Solutions, India Dr. Lamia Bouafif High Engineering Institute of Tunis, Tunisia Prof. Maringanti Hima Bindu North Orissa University, India Dr. Belghini Naouar SIA Lab, Faculty of Science and Technology, Morocco Dr. Fatima Debbat University of Mascara, Algeria
Prof. H.R. Kamath Malwa Institute of Technology, India
Prof. You Xiaoming Shanghai University of Engineering Science, China
Dr. Deacha Puangdownreong SouthEast Asia University, Thailand
Dr. Sanjiv Kumar B. P.S. Mahila Vishwavidyalaya, India
Prof. Jyoti Ohri National Institute of Technology, India
Dr. Muazzam A. Siddiqui King Abdulaziz University, Saudi Arabia
International Journal of Intelligent Systems and Applications(IJISA, ISSN Print: 2074904X, ISSN Online: 20749058) is published monthly by the MECS Publisher, Unit B 13/F PRAT COMMâ€™L BLDG, 1719 PRAT AVENUE, TSIMSHATSUI KLN, Hong Kong, Email: ijisa@mecspress.org, Website: www.mecspress.org. The current and past issues are made available online at www.mecspress.org/ijisa. Opinions expressed in the papers are those of the author(s) and do not necessarily express the opinions of the editors or the MECS publisher. The papers are published as presented and without change, in the interests of timely dissemination. Copyright ÂŠ by MECS Publisher. All rights reserved. No part of the material protected by this copyright notice may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording or by any information storage and retrieval system, without written permission from the copyright owner.
International Journal of Intelligent Systems and Applications (IJISA) ISSN Print: 2074904X, ISSN Online: 20749058 Volume 8, Number 12, December 2016
Contents REGULAR PAPERS Image Superresolution via Divergence Matrix and Automatic Detection of Crossover Dmytro Peleshko, Taras Rak, Ivan Izonin
1
Bezier Curves Satisfiability Model in Enhanced Hopfield Network Mohd Shareduwan M. Kasihmuddin, Mohd Asyraf Mansor, Saratha Sathasivam
9
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks Vu Chien Thang, Nguyen Van Tao
18
A Review of Methods of Instancebased Automatic Image Annotation Morad A. Derakhshan, Vafa B. Maihami
26
A Survey on Speech Enhancement Methodologies Ravi Kumar. K, P.V. Subbaiah
37
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT Yingwei Xiao
46
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems Zhengbing Hu, Vadym Mukhin, Yaroslav Kornaga, Yaroslav Lavrenko, Oleg Barabash, Oksana Herasymenko
57
Empirical Study of Impact of Various Concept Drifts in Data Stream Mining Methods Veena Mittal, Indu Kashyap
65
I.J. Intelligent Systems and Applications, 2016, 12, 18 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.01
Image Superresolution via Divergence Matrix and Automatic Detection of Crossover Dmytro Peleshko Lviv Polytechnic National University / Department of Publishing Information Technologies, Lviv, 79013, Ukraine Email: dpeleshko@gmail.com
Taras Rak Lviv State University of Life Safety, Lviv, 79013, Ukraine Email: rak.taras74@gmail.co m
Ivan Izonin Lviv Polytechnic National University / Department of Publishing Information Technologies, Lviv, 79013, Ukraine Email: ivanizonin@gmail.co m
Abstract—The paper describes the image superresolution method with aggregate divergence matrix and automatic detection of crossover. Formulation of the problem, building extreme optimization task and its solution for solving the automation determination of the crossover coefficient is presented. Different ways for building oversampling images algorith ms based on the proposed method are shows. Based on practical experiments shows the effectiveness of the procedure of automatically the determination of the crossover coefficient. Experimentally established the effectiveness of the procedures oversampling images at high zoo m resolution by the developed method. Index Terms—Superresolution, similarity measure, crossover operations, automatic detection, aggregate divergence matrix.
ensure qualitative results oversampling at high zoo m for images that define the fluctuating intensity function. Images obtained using such methods can be used in many applications, concerning medicine, criminology, remote sensing, and more. The aim of this art icle is to modify the method of providing superresolution in the case of t wo input images by automating determining the factor o f crossover operation at each successive stage zooming. For this, in the work describe the task of automatically determin ing the factor of crossover operation which is solved by solving an optimization problem by criteria built on the continuum of degrees of similarity. The advantages of this approach compared to existing obvious: the possibility of full auto mation of the method; the selection of the correct values of crossover using solutions of the optimization problem, for efficient work of the method, and so on.
I. INT RODUCT ION Develop ment of new and imp rovement of existing methods and tools for superresolution digital images that provide effective results process of oversampling image is a challenging problem due to a number of reasons. On the one hand  the prevalence of this type of problems in modern intellectual systems of image processing, and the other  the lack of effective methods for solving this problem in the case of two input images of a scene. Most modern methods to increase the resolution focused on the handling of images with s mooth function of intensity, and their adaptation to solve the problem to ensure effective results at high magnification, especially in the case of fluctuation function is far fro m a co mplete solution. The complexity of this problem beco me deep throughout needed to preserve image edges sharpening with the intensity fluctuation function and reduce artifacts and distortions that occur in the input image processing. Therefore, the urgent task is to develop methods and the development of image superresolution that would Copyright © 2016 MECS
II. A NALYSIS OF PREVIOUS ST UDIES Several classical methods (bilinear, bicubic interpolations) of images resolution increase are characterized by high speed performance. Ho wever, they include a number of artifacts that impose restrictions on the practical applicat ion of these methods. In addition, they operate just by one image. In some cases, it can be considered as disadvantage. In particular, the presence of several samples o f one scene that are aligned relative to each other in partial pixel value carry additional informat ion that could be used for increasing the resolution as well as for improvement of the original sample quality. One of the possible solutions to this problem is the use of Super resolution reconstruction methods, or superresolution. In general, the superresolution technology consists of two classes of methods [14], including: optical (OSR) and geo metric (GSR) methods of superresolution. The first class (OSR) – these are methods focused on the I.J. Intelligent Systems and Applications, 2016, 12, 18
2
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
hardware imp lementation. While the second class (GSR) – these are methods, algorithmic imp lementation of which is based on the processing of similar images or their frag ments sets. The idea of the geometric methods class of superresolution – increasing of image or set of images on the basis of one or more samples of low resolution, for wh ich pixel or subpixel shift is typical. When we talk about super resolution methods based on several samples of low resolution of one scene, the important point here is the presence of pixel or sub pixel shift between them. It is the carrier of the additional informat ion that will provide the opportunity to obtain image with high information content. The rapid development of this class methods is characterized by the use of different tools to solve the set problem. Accordingly, developed a nu mber of these methods classifications is developed [15], including the quantity of low resolution images, which are used in the work process [16] or concerning the area where these methods work [17], or concerning the actual method of the original sample reconstruction, and so on. The most detailed classification of superresolution methods, in our opinion, is provided in [18]. According to it, two methods groups of images superresolution of the spatial do main are considered: a) classical  the methods which work with several samples of low resolution and b) on the basis of one image processing (Fig. 2). The authors in [18] define the following nu mber o f classical methods: methods of iterative back project ion, iterative adaptive filtering, direct methods, methods based on the theory of projections on convex sets and probabilistic methods. It should be added that this classification is inco mplete, since in recent years’ methods that combine various mathematical apparatus of the above mentioned groups and form a group of hybrid methods have developed [19, 20, 21, 22, 24]. Methods use of iterative back project ing [23] can cause such artifacts in the resulting image as effect of bell and aliasing. In addit ion, iterative process of error reconstruction min imization can converge to a few results. This causes the existence of mu ltip le solutions; the algorith m may oscillate between them or converge to one of them. Methods of iterative adaptive filtering are designed to handle video streams. Their main purpose is the unknown variables estimation that is why they mainly use a Kalman filter. Methods group [25, 28, 27] on the basis of projections theory on convex sets is based on iterative finding of acceptable region element which is defined by the intersection of a nu mber of convex constraints, starting fro m an arbit rary point. The main problem here is the definit ion of p rojections, wh ich may occur as a not triv ial task. In addition, there may be several solutions of these methods. Another subgroup of methods [26]  statistical solves
Copyright © 2016 MECS
the problem of super  resolution stochastically by close to optimal image reconstruction. The resulting image, and underframes movement are treated as stochastic variables. These methods formulate the problem of super resolution as an inverse problem, which is sensitive to a variety of art ifacts. In addition, if the number of lo wresolution images which are used by the method is less than the rate of resolution increase of the original image it is again reduced to incorrect problem. In other words, methods in this class also do not provide the uniqueness of the problem solution.
III. IMAGES PREPROCESSING As in our previous work [2], input data of the developed method are two same resolution image: j 1..l
I1 c1 i , j i 1..h
and
j 1..l
, I 2 c2 i , j i 1..h
(1)
where ci,j  the function of pixel intensity with coordinates (i, j). The work proposed method requires preliminary processing on pairs of images, which determined execution following procedures. 1. Normalizat ion of images I1 and I2 according to the expressions:
1 ci , j ci , j max ci , j min ci , j , i1; h; 2 ij1;1;hl ; j1;l ci , j Kci , j ,
(2)
(3)
1
where K max ci , j . As a result, we get: i1;h; j1;l i 1; h , j 1;l :ci , j 0. 2.
Construction of the new vector
c i (using
normalized input images (2)) fro m the corresponding lines (columns) using singlepoint crossover operation:
c i kс1 i 1 k с2 i ,
(4)
where k  crossover operations coefficient; c1 i , c2 i – dimension vectors l, elements of which are matrix ro ws I1 and I2 respectively. Use crossover operations to build new lines of image, shown in Fig. 1. The paper used the first case of variant b).
I.J. Intelligent Systems and Applications, 2016, 12, 18
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
3
Fig.1. Using crossover operation for the synthesis of new line of image from two input
As a result, we obtain a new matrix: j 1..l
I c i c i , j i 1..h i 1..h
(5)
If the horizontal direction increase, matrix I will look like:
I c j j 1..l
(6)
where c j vectors constructed according to (4) for the
V. DEFINING VECT ORSFEAT URES BASED ON M OOREPENROUZE PSEUDOROT AT ION OF DIVERGENCE M AT RIX The next step of the proposed method is to search vectorsfeatures, which act as the image I characteristics of each line (column). To solve the problem of characteristic vectors yi building, which will perform as the characteristics of each line (or column) image C, we consider the equation:
i yi сi ,
(11)
corresponding columns normalized input images I1 and I2 respectively.
where сi сi , j  j 1..l ; yi = (yi ,1 , …, yi,l) – l
IV. DIVERGENCE M AT RIX CONST RUCT ION
measurable vector of image С characteristic values for the ith line. In general, the characteristic vector yi (7) is defined as:
Aggregate divergence matrix i , that is the basis of the superresolution method, in the case of two input images, is constructed as follows [2]:
i 1; h : i Ai ci
T
... ci , l
ci ,1 ... ci ,1 1 j Ai ... ... ... ; c ci , x i, j j x 1 max ci , j i1, h ci ,l ... ci ,l j1, h 1
(7)
(8)
c j ,1 ... c j ,1 Aj ... ... ... ; max ci , j i1, h c j , h ... c j , h j1, h 1
Copyright © 2016 MECS
Matrix і is singular (det(і ) = 0), so the inverse matrix i1 does not exist. Therefore, to solve the problem (12) according to the theory on the minimizat ion of the residual error xi  і yi 2 of the linear system [3] the following way of vector yi defining is proposed:
yi i сi (1 i i )ri ,
(13)
(pseudorotation to і matrix [1, 3]); (1 i i ) –
T
... c j Aj , h
(12)
where i – generalized inverse matrix MoorePenrose
In applying the procedures for changing the resolution on rows of images I aggregate divergence matrix defined by (7)  (8) will look like:
j 1; l : j c j
yi i1сi .
(9)
1 i c j ,i cx , j (10) i x 1
projecting operator on the kernel і ; ri – random vector of dimension l. The first term у (13) acts as pseudorotated solution, and the second is the solution of ho mogeneous system і yi = 0. Provided through (13) method of defining vector characteristics of ith element of overlapping is possible as according to [3] matrix i i stoops to be бути inverse. MoorePenrose matrix i is defined according to matrix singular distribution і in the following way [3]:
i Vi iUiT ,
(14)
I.J. Intelligent Systems and Applications, 2016, 12, 18
4
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
where Ui , Vi – unitary matrices of order l×l of matrix singular distribution і;
error:
i
– matrix o f order l×l, wh ich
is pseudorotated to the diagonal matrix i of singular matrix d istribution і . As matrix i is also inverse, then matrix
VI. A UT OMATICALLY DETERMINING T HE COEFFICIENT OF CROSSOVER OPERAT ION IN T ASKS DIGIT AL IMAGES OVERSAMPLING
i
is received fro m i by replacing all nonzero
singular values i,q (i,1 i,2 … i,l 0) to respectively conversed to them 1/i, q . In the iteration process of finding according to yij 1 with (12) random vector
ri j 1 = xi  і yi j l , where – lnorm.
To develop the automated determination of k use approach to building optimizat ion criterion based on a continuum of Syomkin similarity measures [5]:
ri j 1 was defined by residual
1
K I  I ет K , I ет  I K , I , I ет , ; 1 , , 2 where – the measure of pro ximity of neighboring objects by Kolmogorov [6]; {+, 1, 0, 1, } – element of set plural ordered that defines one of the used similarity measures. For example, when = 0, = 1 we obtain the plural of Kulchynskiy K0,1 [7]. We proposed criteria feature of optimization task: 1 K k Ki , k K1, 1 k , 6 i, 1,0,1,
(15)
where K0,1 , K0,0 , K0, 1 , K0, , K0, , K1, 1 – Ku lchynska appropriate plural [7], Otiai [8], Sorensen [9] BraunBlanquet [10] Shy mkevychaSimpson [11, 12], Jacquard [13]. In the event of I(2) and reference the twice en larged image I m 2 these plurals as described in [13] at the fixed value of the coefficient k are written as follows:
(16)
2 I I ; K I , I K I , I I I I I ; I I I I 2 I I K I , I ; K I , I ; I I 2 I I I I I I 2 I I I I I I K I , I ;K , I , I I I I I I I 2
2
I 2 I ет 2
ет
0,0
2
2
2
0,
2 ет
2
2
2 ет
0,1
2
2
2
1, 1
2 ет
2
0,
k0,1
0, 1
2 ет
2 ет
(17)
It is worth paying attention that experimentally established that the optimization problem enough to solve just one of the stages consistent increase in the oversampling image task. The resulting value will be optimal or close to optimal in all the other stages of solving the task.
2 ет
2 ет
2
where  determine the dimension of the plural operator. Choice of these coefficient for build criterial features (7) conditioned by measures topological equivalence. Then the optimization problem to select optimal values of crossover operation (opt(k)) can be written as:
opt k arg max K k .
2
2 ет
2 ет
2
2 ет
2 ет
2
2 ет
2
2 ет
2 ет
VII. A LGORIT HMIC IMPLEMENT AT ION OF M ET HOD Solution of (12) with the operator of (7) or (10) is used to solve the problem of t woframe image superresolution in the case of crossover transactions. Since the input in this case there are two images І1 and І2 , there are at least three solutions of the problem oversamp ling. The first two are similar to [1, 2] consists in the build ing enhanced images added to the original matrix and І1 or І2 vector сi + yi position in line (or colu mn). Here сi – the function of intensity of the input image (i.e., the original, not normalized by (2) and (3) of significance). A third solution is the synthesis of matrix rows and colu mns, enhanced by the expression:
c3 i
Copyright © 2016 MECS
2
2
2
2
2 ет
2 ет
2
2
2 ет
2
2 ет
2 ет
2 ет
2 ет
c1 i c2 i 2
y1 i , i 1..l
(18)
I.J. Intelligent Systems and Applications, 2016, 12, 18
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
SR increase procedure similar to the procedure described in [1, 2], consists of two consecutive parts, which oversampling image carried in the vertical and horizontal d irections, respectively. Ordering action of each part of the algorithm can be arbitrary. The algorith m for the image superresolution problem solving involves the following steps:
5
The first three parts consistently applied to all the ro ws of I to increase the size of a given image height. Only then enlarged image matrix I1 m , I 2 m or I m built by (15) or like according to [2]. Further procedure of increase implemented on larger image I1 2 , I 2 2 or I 2 until the variable m reaches targeted coefficient increase.
1) construction mutated input vector fro m vectors corresponding two input images based on (4). 2) construction aggregate divergence matrix (7) and (9). 3) calculation of characteristic vectors y i square matrices built on relat ionships (7 o r 9), on the iterat ive procedure (13). 4) building enhanced image by adding to the orig inal matrix I1 or I2 characteristic vectors like [2], or synthesis increased images according to (15).
VIII. RESULT S OF PRACT ICAL EXPERIMENT S In order to assess the effectiveness of oversampling image by the developed method was conducted a series of practical experiments. A couple of input different qualities images, which used for simu lation of the method shown in Fig.2. The crossover coefficient: k = 0.7. This value obtained by solving the optimizat ion task used in the work to automatically determine k (Fig. 3.)
a)
b)
Fig.2. A pair of input 2byte, grayscale image, resolution, dimension – 231199 pixels: а) I 1 ; b) I 2
0,78 0,77
K
0,76
0,75 0,74 0,73
0,72 0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
k Fig.3. Dependence of the optimization criterion from values of operation crossover
Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 18
6
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
of PSNR (ie the growth with smaller gradient). A significant increase PSNR in cases coefficient resolution change grater may be the main argu ment for practical using oversampling procedures, based on the proposed method. Co mparing the results of the proposed method and existing (in the case of one input image [1]) shows the following results (Fig.5).
IX. DISCUSSION Fig. 4 shows PSNR depends, resulting oversampling of images І1 , І2 and the synthesis of image I m according to (15). As shown in Fig. 4 oversamp ling best results obtained at the synthesized image І(m). It is worth noting about PSNR behavior change. Three areas PSNR determine the intervals drop, growth and satiation values 45
PSNR
40 35 30 25 20
0
2
4
6
8
10
12
14
16
18
Rate of increase resolution I1
I2
I
Fig.4. Dependence PSNR from coefficient of increase for images I1 m , I 2 m and I m obtained by method
42
PSNR
41
40 39 38 37
0
2
4
6
8
10
12
14
16
18
Rate of increase resolution, m existing method
developed method
Fig.5. Comparison of the work developed and existing methods based on PSNR when ch anging m
Fro m the Fig. 4 we can say that for small coefficient values of increase (m <10) using the proposed method is not justified since PSNR values for existing methods are considerably higher. The value of PSNR developed method starting to exceed the value of PSNR for existing only when m> 10. Obviously, the exact value of the inflection point m = 10 in the co mparison of the effectiveness of the methods obtained for a g iven image І1 and І2 . However, experiments with other images in general, confirmed that for other images this point will be close to exactly value of 10.
[2]
[3]
REFERENCES [1]
Iu. M . Rashkevych, I.V. Izonin, D.D. Peleshko, I.O. M alets ―Zmina rozdilnoi zdatnosti zobrazhennia zasobom
Copyright © 2016 MECS
[4]
psevdoobertannia vyrodzhenoho matrychnoho operatora vidnosnykh symetrychnykh mir konverhentsii‖, in Bulletin of the Lviv Polytechnic National University, Computer Science and Information Technology, № 826, pp. 259–266, 2015 (in Ukrainian) Dmytro Peleshko, Taras Rak, M arta Peleshko, Ivan Izonin, and Danylo Batyuk ―Twoframes image superresolution based on the aggregate divergence matrix‖, Data stream Mining & Processing: proc. of the 1st international scien. and techn. conf., 23–27 August 2016, Lviv, Ukraine: Lviv Polytechnic Publishing House, 2016, pp. 235–238. doi: 10.1109/DSM P.2016.7583548 R. Penrose ―A generalized inverse for matrices‖, in Proceedings of the Cambridge Philosophical Society 51, pp. 406413, 1955. Yevgeniy V. Bodyanskiy, Elena A. Vynokurova, Artem I. Dolotov "SelfLearning сascade spiking neural network
I.J. Intelligent Systems and Applications, 2016, 12, 18
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14] [15]
[16]
[17]
[18]
[19]
[20]
for fuzzy clustering based on group method of data handling" in Journal of Automation and Information Sciences, vol 45, № 3, pp. 2333, 2013. Semkyn, B. Y. and Dvoichenkov, B. Y. (1973) "Ob tkvyvalentnosty mer skhodstva y razlychyia" Yssledovanye system. T. 1: Analyz slozhnykh system, Vol 1, pp. 4576. (in Russian) Kliushyn, D.A. and Petunyn, Iu.I. (2003) ―Neparametrycheskyi kryteryi ekvyvalentnosty heneralnykh sovokupnostei, osnovannyi na mere blyzosty mezhdu vyborkamy‖, in Ukrainskyi matematychnyi zhurnal [Ukrainian M athematical Journal], V.5, № 2, pp.147– 163. (in Ukrainian) Kulczinsky, S. (1927) ―Zespoly, rуslin w Pienach‖, Bull. intern. acad. polon. sci. lett. Cl. sci. math. natur., Ser. B, S.2. pp. 241—272. Ochiai, A. (1957) ―Zoogeographical studies on the soleoid fishes found Japan and its neighboring regions. II‖, Bull. Jap. Soc. sci. Fish., Vol. 22, № 9, pp. 526—530. Sorensen, T. (1948) ―A method of establishing groups of equal amplitude in plant sociology based on similarity of species content‖, Kongelige Danske Videnskabernes Selskab. Biol. Krifter, № 4, pp. 134. Braun–Blanquet, J. (1951) ―Pflanzensoziologie:Grundzge der Vegetationskunde», Berlin: Springer–Verlag WIEN GMBN, 530 p. Simpson G.G. (1947) ―Holarctic mammalian faunas and continental relationship during the Cenozoic‖, Bull. Geol. Sci. America, Vol. 58, pp. 613688. Szymkiewicz, D. (1934) ―Une contribution statistique a la geographie floristique‖, Acta Soc. Bot. Polon, Vol. 34, № 3, pp. 249—265. Sѐmkyn, B. Y. (2007) ―Kolychestvennыe pokazately dlia otsenky odnostoronnykh florystycheskykh sviazei, predlozhennыkh B. A. Yurtsevym‖, Bot. Zh, V. 92, pp. 114–127. Superresolution. Electronic resource: http://www.wikiwand.com/en/Superresolution. K. Nasrollahi and T.B. Moeslund ―Superresolution: a comprehensive survey‖, in Machine Vision and Applications, Vol. 25, № 6, pp. 1423–1468, 2014. I. Izonin, R. Tkachenko, D. Peleshko, T. Rak and D. Batyuk, "Learningbased image superresolution using weight coefficients of synaptic connections," Scientific and Technical Conference "Computer Sciences and Information Technologies" (CSIT), 2015 Xth International, Lviv, 2015, pp. 2529. doi: 10.1109/STCCSIT.2015.7325423 D. Glasner, S. Bagon and M . Irani ―SuperResolution from a Single Image‖ in Computer Vision: proc. of 12th intern. conf., Kyoto, 27 Sept. – 4 Oct. 2009, IEEE, pp. 349 – 356. K. Nasrollahi and T.B. Moeslund ―Superresolution: a comprehensive survey‖, in Machine Vision and Applications, Vol. 25, № 6, pp. 1423–1468, 2014. M . Elad and A. ―Feuer Restoration of a single superresolution image from several blurred, noisy and downsampled measured images‖ in IEEE Transactions on Image Processing, Vol. 6, № 12, pp. 1646–1658, 1997 A. J. Patti, Y. Altunbasak ―Artifact reduction for set theoretic super resolution image reconstruction with edge adaptive constraints and higherorder interpolation‖, in IEEE Transaction on Image Processing, Vol. 10, № 1, pp. 179–186, 2001
Copyright © 2016 MECS
7
[21] A. J. Patti, M . Sezan, and A. M . Tekalp ―Robust methods for high quality stills from interlaced video in the presence of dominant motion‖, in IEEE Transactions on Circuits and Systems for Video Technology, Vol. 7, № 2, pp. 328342, 1997 [22] Wanqiang Shen, Lincong Fang, Xiang Chen, Honglin Xu ―Projection onto Convex Sets M ethod in Spacefrequency Domain for Super Resolution‖, in Journal of computers, Vol. 9, № 8, pp. 1959–1966, 2014 [23] Y. W. Tai, W. S. Tong, and C. K. Tang ―Perceptuallyinspired and edgedirected color image superresolution‖, in Computer Vision and Pattern Recognition: proc. of intern. conf, New York, 17 – 22 June 2006, Los Alamitos: IEEE CS, 2006, Vol. 2., pp . 19481955. [24] Zhifei Tang, Deng M ., Chuangbai Xiao and Jing Yu ―Projection onto convex sets superresolution image reconstruction based on wavelet bicubic interpolation‖, in Electronic and Mechanical Engineering and Information Technology (EMEIT): proc. of intern. conf., Harbin, 1214 Aug. 2011, IEEE Press, 2011, Vol. 2, pp. 351  354. [25] Frederick W. Wheeler, Ralph T. Hoctor and Eamon B. Barrett ―SuperResolution Image Synthesis using Projections onto Convex Sets in the Frequency Domain‖, in Computational Imaging, Electronic Imaging Symposium, : proc. of intern. conf., San Jose, January 2005, Vol. 5, pp. 479490. [26] S. Cain, R. C. Hardie, and E. E. Armstrong ―Restoration of aliased video sequences via a maximumlikelihood approach‖, in Passive Sensors: proc. of nat. infrared inform. symp., M onterey, M ar. 1996, pp. 377–390. [27] Wanqiang Shen, Lincong Fang, Xiang Chen, Honglin Xu ―Projection onto convex sets method in spacefrequency domain for super resolution‖, in Journal of computers, Vol. 9, № 8, pp. 1959–1966, 2014. [28] Zhifei Tang, Deng M ., Chuangbai Xiao, Jing Yu ―Projection onto convex sets super–resolution image reconstruction based on wavelet bi–cubic interpolation‖ in Electronic and Mechanical Engineering and Information Technology (EMEIT): proc. of intern. conf., Harbin, Vol. 2, P. 351 – 354, 2011.
Authors’ Profiles Dmytro Peleshko, Dr. Sc., Professor at Lviv Polytechnic National University, Lviv, Ukraine. He has published more than 100 papers in international and national scientific issues and journals and he is the author of several monographs. The main research interests include image and video processing for the system of artificial intelligence
Rak Taras, Dr. Sc., an Associate Professor and vicerector on scientific and research work at Lviv State University of Life Safety. He received a Doctor of Sciences degree in Information Technology. He has published more than 100 papers in international and national scientific issues. The main research interests include information technology, decisionmaking system, control systems for emergencies.
I.J. Intelligent Systems and Applications, 2016, 12, 18
8
Image Superresolution via Divergence M atrix and Automatic Detection of Crossover
Ivan Izonin, Ph.D, teacher assistent at Lviv Polytechnic National University, Lviv, Ukraine. He has published more than 50 papers in international and national scientific issues and journals and he is the author of tutorial. The main research interests include image and video processing for the system of artificial intelligence.
How to cite this paper: Dmytro Peleshko, Taras Rak, Ivan Izonin, "Image Superresolution via Divergence M atrix and Automatic Detection of Crossover", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.18, 2016. DOI: 10.5815/ijisa.2016.12.01
Copyright ÂŠ 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 18
I.J. Intelligent Systems and Applications, 2016, 12, 917 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.02
Bezier Curves Satisfiability Model in Enhanced Hopfield Network Mohd Shareduwan M. Kasihmuddin School of Mathematical Sciences, Universiti Sains Malaysia, 11800 USM, Penang Malaysia Email: iwanmaidin@gmail.co m
Mohd Asyraf Mansor School of Mathematical Sciences, Universiti Sains Malaysia, 11800, Pulau Pinang, Malaysia Email: asyrafalvez@live.com
Saratha Sathasivam School of Mathematical Sciences, Universiti Sains Malaysia, 11800 USM, Penang Malaysia Email: saratha@usm.my
Abstract—Bezier cu rve is one of the most pragmatic curves that has vast application in co mputer aided geometry design. Unlike other normal curves, any Bezier curve model must follow the properties of Bezier curve. In our paper, we proposed the reconstruction of Bezier models by imp le menting satisfiability problem in Hopfield neural network as Bezier properties verification technique. We represent our logic construction to 2satisfiability (2SAT) clauses in order to represent the properties of the Bezier curve model. The developed Bezier model will be integrated with Hopfield neural network in order to detect the existence of any nonBezier curve. Microsoft Visual C++ 2013 is used as a platform for training, testing and validating of our proposed design. Hence, the performance of our proposed technique is evaluated based on global Bezier model and computation time. It has been observed that most of the model produced by HNN2SAT are Bezier curve models. Index Terms—Bezier curve, Hopfield network, 2satisfiability, Logic programming, Wan Abdullah‟s method.
I. INT RODUCT ION Co mputer Aided Geo met ric Design (CA GD) has emerged as one of the most eminent fields in co mputer graphic, numerical co mputation, and geometrical studies. The renaissance of CA GD is inaugurated by the various form of Be zier curves that has attracted a prolific nu mber of research [1]. Hence, Bezier curves are one of the most intensively used in CA GD [2]. Popularized by Bezier [3], Bezier curves are used to construct smooth curves at any scale by considering its own properties to be applied in the design. The main impetus of this research is to reconstruct the Bezier curves according to the correct properties by using neural network approach. Hence, the Bezier properties verification process requires a robust and stable algorithm in order to construct more co mp lex Copyright © 2016 MECS
Bezier curves. Thus, we choose to translate the Bezier curves properties into 2Sat isfiability problem and integrate it in Hopfield neural network by using logic programming (HNN2SAT). Basically, 2Sat isfiability (2SAT) is the prominent counterpart of the Boolean satisfiability (SAT) optimization problem, that is denoted in Conjunctive Normal Form (CNF) form [4, 8]. W ithout a doubt, there have been a lot of applications of 2Satisfiability such as the previous work by Femer [5], Papadimit riou [6], Patreschi and Simeone [7]. These related wo rks emphasized on the 2Satisfiab ility as an optimization problem. On the separate note, we can transform any real dataset or mathematical p roperties into 2SAT form with the assistance of logic programming [9]. In this paper, we emphasized on the Bezier curve properties as 2SAT co mbinatorial optimizat ion problems in logic programming. The Hopfield neural network p lays an important role in the field of art ificial intelligence and mathemat ical computation. Recurrent Hopfield neural networks are principally dynamical schemes that feedback signals to themselves. The network was inaugurated by Hopfield and Tank [10]. One of the interesting features is the network possess a dynamical system with stable states with each own basin and attraction [11]. Moreover, the Hopfield neural network min imizes Lyapunov energy because of physical spin of the neuron states . On top of that, the network produced global output by minimizing the network energy. Gad i Pinkas and Wan Abdullah [9, 12] demarcated a bidirect ional mapping between logic and energy function of a symmetric neural network. Besides, both related works are the building blocks for a corresponding logic program. The work of Sathasivam [13] presented that the optimized recurrent Hopfield network could be possibly used to do logic programming. Above all, the ability of learn ing by using Hopfield network is the main priority in the reconstruction of the Bezier curves. The memory will be stored in Hopfield‟s brain as content addressable memory (CAM) [14]. Hence, we can reconstruct the correct Bezier curves by retrieving I.J. Intelligent Systems and Applications, 2016, 12, 917
10
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
the stored memory based on the properties. Logic Programming refers to an auspicious field, and widely used to resolve numerous constraint optimizat ion problem [15]. Additionally, the logic programming can be demarcated as an optimization problem [13, 16]. In this paper, the Bezier properties will be represented by the clauses in 2SAT. After that, 2SAT problem will be translated into logic programming integrated with Hopfield neural network. The impetus of our study is the wellknown Wan Abdullah‟s technique [9, 12]. Wan Abdullah presented a model of doing logic programming in Hopfield network by opt for the Horn clauses as the problem [12, 17]. Theoretically, the Hopfield network can min imize the logical inconsistency in any order of logic p rograms [18]. Thus, we develop a network called HNN2SAT to be tested with the Bezier satisfiability problem. The implementation of Hopfield neural network, 2Satisfiab ility and logic programming as a hybrid model (HNN2SAT) in the reconstruction of various Bezier curves is the main contribution of the study. The rest of the paper is organized as follows. Section II contains the fundamental conceptual background of satisfiability problems, especially the 2satisfiability (2SAT) problem. In Section III, we d iscuss the core theory of Bezier satisfiability and the important properties of Bezier curves. In addition, we emphasize the different type of Bezier curves being used in our proposed Bezier satisfiability problem. Section IV accentuates the neuro logic parad ig m, consists of Hopfield neural network and logic programming. This will include the implementation of Hopfield network and logic programming in reconstructing the Bezier curves. Section V describes the theory imp lementation for our proposed algorithm in doing Bezier satisfiability. Section VI presents the result and analysis. In this section, the discussion will point on the trend of global Bezier curve and the running time recorded for d ifferent type of Bezier curves. Sect ion VII then highlights the future work and encloses the conclusion of this work.
II. SAT ISFIABILIT Y (SAT) PROBLEM Satisfiability (SAT) is a significant problem in computer science and mathematics [4]. SAT problem helps researcher to deal with constraint optimizat ion problem such as circuit and pattern reconstruction. One way to learn v ia SAT is by embedding our required informat ion inside and SAT problem and solve it optimally. In Bezier reconstruction, learning and verify ing is Bezier curve model work identically with SAT problem. In general, a Boolean formu la is satisfiable if there exists an assignment of values true and false that makes the entire expression true [5]. The easiest way to solve SAT problem is by utilizing exhaustive search method, where SAT will try out every possible truth assignment. For examp le, g iven a problem size n , there will n
be 2 such assignments and l literals to set for each assignment [13, 27]. At this point, this method involves Copyright © 2016 MECS
O l.2n operations [28] and it was proven by many that
SAT is an NPcomplete problem. SAT problem normally represented in Boolean variables or expressions in conjunctive normal form (CNF). CNF is defined as conjunction of clauses, where the clauses are disjunction of literal [35]. Literal is a variable or its negation. For example:
x1 x2 x2 x3 x5 x1 x4
(1)
Based on (1), x1 , x2 , x3 , x4 are Boolean variables to be assigned, means negations (logical NOT), means negations (logical OR), means negations (logical AND). We can satisfy formula (1) by taking x1 true, x2 false, x3 false, x4 true . However, if any formula is not satisfiable, it will be termed as unsatisfiable. In Bezier model reconstruction, we integrate the foundation of satisfiability in order to obtain the correct Bezier model. The literals in every clause will be represented the properties of the Bezier curve model. A. 2SAT 2SAT is a subset of SAT problem. It is a classical NP problem that determine the satisfiability o f sets of clauses with at most two literals per c lause (2CNF formu las) [8]. Besides, it is popular and h ighly regard problem for general Boolean satisfiability which can involve constraints on two variables [31]. In addition, the variables can allow two possibilities for the value of each variable. 2SAT problem can be expressed as 2CNF (2Conjunctive Normal Form). Randomized 2SAT problem is considered as NP problem or nondetermin istic problem [6]. The three fundamental co mponents of 2SAT are summarized as follows: 1. 2. 3.
A set of m variables, x1 , x2 ,......, xm A set of literals. A literal is a variab le or a negation of a variable. A set of n distinct clauses: C1 , C2 ........Cn . Each clause consists of only literals combined by just logical OR ( ). Each clause must consist of 2 variables.
The Boolean values are 1, 1 . Researchers have replaced F and T in the neural networks by 1 and 1, respectively to emphasized false and true. Since each variable can take only two values, a statement with n variables requires a table with 2n rows. The goal of the 2SAT problem is to determine whether there exits an assignment of truth values to variables that makes the following Boolean formula P satisfiable. n
P Ci i 1
(2)
Where is a logical AND connector. Ci is a clausal I.J. Intelligent Systems and Applications, 2016, 12, 917
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
11
form of DNF with 2 variables. Each clause in 2SAT has the following form n
Ci xi , yi i 1
(3)
xi ki , ki and yi ri , ri ki and ri are negations
of the literals.
III. BEZIER M ODEL SAT ISFIABILIT Y Popularized by French engineer, Pierre Bezier, Bezier curve is a parametric curve that frequently used in computer graphics and design automobile bodies . Currently, Bezier curves are extensively used to model smooth curves and other applications [20]. A Bezier curve is defined by a set of control points P0 through Pn , where n is the o rder of the curve. In essence, the first and the last control point are always the end point of the curve. The other point between the first control point to the last do not lie on the curve [21]. Since the curve is co mpletely enclosed in the convex hull of its control points, the points can be manipulated graphically. A. Properties of Bezier Curve Bezier curve is a parametric curve that uses the Bernstein polynomials as a basis. Generally, a Bezier curve of degree n (Order n+1) is represented by
B t
n
b
i,n
t Pi
, 0 t 1
(4)
Fig.1. Basis function for cubic Bezier
2) Geometric point property: (i) The first and the last control points are the endpoints of the curve. Moreover, b0 B 0 and
bn B 1 . (ii) The curve is a tangent to the control polygon at the endpoints [21]. This can be verified by taking first derivatives of a Bezier curve B ' t
In
dB t dt
detail,
n 1
n
n n i bi ,n t t i 1 t , n 0,1, 2,3,....., n i
i 1
Pi bi ,n 1 t , 0 t 1 (6)
i 0
we
acquire
P ' 0 n P1 P0
and
B ' 1 n Pn Pn1 . The above equation can further simplified by setting Pi Pi 1 Pi :
i 0
B ' t n
where
P
n 1
P B
i i , n 1
t ,
0 t 1
(7)
i 0
(5)
The shape of the curve can be determined by using the coefficient Pi which are the control points and Bi ,n t are the basis of the function [21]. Lines can be drawn between the control points of the curve in order to form control polygon. Bezier curve possessed the following properties
3) Convex hull property: A domain D is considered as a convex if any two points P1 and P2 in the domain, the segment P1 P2 is entirely embodied in the domain D [23]. The convex hull of a set of points P is the boundary of the smallest convex domain containing P. The entire curve is contained within the convex hull of the control point.
1) Geometry invariance property: Partit ion of unity of the Bernstein polynomial satisfy any of the Bezier curve although the curve undergoes translation and rotation of its control point [22]. The basis function of any Bezier curve can be represented by using the following graph: Fig.2. Convex hull built from the control point
B. Type of Bezier Curve A Bezier curve is defined based on their control points. For our studies, we will examine the following curve Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 917
12
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
1) Linear Bezier curves Given points P0 and P1 , a linear Bezier curve is a straight line between those two points. The curve can be formulated as followed
B t 1 t P0 tP1
(8)
Where 0 t 1 .The equation normally similar to linear interpolation.
Fig.5. Cubic Bezier curve
C. BezierSAT T able 1. Bezier2SAT clause representation Clause
Properties
Descriptions Linear Bezier
C11 1 t
Fig.3. Linear Bezier curve
C12 t
2) Quadratic Bezier curves
Total clauses: 2
Quadratic Bezier curve is a curve that traced by the function B t given points P0 , P1 and P2 . The curve can
Quadratic Bezier
C11 1 t
be formulated as followed:
B t 1 t P0 2 1 t tP1 t P2 2
2
Partition of unity
(9) C1
Where 0 t 1 . The equation is interpreted as the linear interpolant of corresponding points on the linear Bezier curves from P0 to P1 and from P1 to P2 [24].
n
B
i,n
t 1
2
C12 2 1 t C13 t 2 Total clauses: 3
i 0
Quadratic Bezier
C11 1 t
3
C12 3 1 t
2
C13 3 1 t t 2 C14 t 3 Total clauses: 4
Fig.4. Quadratic Bezier curve
C2
3) Cubic Bezier curves
Geometric point where P are the control points of the curve. The curve must touch the first and the endpoint.
Cubic Bezier curve is a curve that traced by the function B t given points P0 , P1 , P2 and P3 . Cubic
3
2
Where 0 t 1 . The curve will not pass through P1 or P2 because these points only to provide directional informat ion. The cubic Bezier curve can be defined as a linear combination of two quadratic Bezier curves [25].
Copyright © 2016 MECS
C22 P1 C23 P2 C24 P3 C2i Pi 1 Total clause : i 1
P0 , P1, P2 , P3 ,....., Pi DC
Bezier curve can be formulated as followed:
B t 1 t P0 3 1 t tP1 3 1 t t 2 P1 t 3 P3 (10)
C21 P0
C3
Convex hull property
Where DC is the domain of the convex hull Total clause : i 1
The properties of Bezier curve with different orders can be represented by using logic. In this case, 2SAT logic is utilized to imp lant the important properties of Bezier curve. If any of the random curve followed all the important properties of Bezier curve [25], the curve model is considered as Satisfiab le (It is a Bezier curve). On contrary, if any of the curve model does not abide with properties, 2SAT logic will consider the curve as I.J. Intelligent Systems and Applications, 2016, 12, 917
Bezier Curves Satisfiability M odel in Enhanced Hop field Network
unsatisfiable (It is not a Bezier curve) [26]. The task of representing the important properties of Bezier curve in 2SAT will be used in training phase of neural network. It was proven by previous researchers that the hybridizat ion between 2SAT logic and neural netwo rk is ab le to solve constraint optimization problem such as pattern reconstruction. Table 1 shows the representation of the properties of the Bezier curve and 2SAT clause during training phase.
Hopfield network work asynchronously with each neuron updating their state determin istically. The system consists of N formal neurons, each is described by an Ising variable. Neurons are bipolar Si 1, 1 obeying the dynamics Si sgn hi where the local field hi . The connection model can be generalized to include higher order connection. This changes the field to
hi
W S
A. Hopfield Model In Bezier curve reconstruction, we select the Hopfield neural network because it is well distributed. Hence, Hopfield neural network is easier to be integrated with any paradigms to solve satisfiability problem [31]. Technically, we incorporate the 2SAT with 3 clauses in Hopfield neural network so that we can relate it with the Bezier cu rve properties. The correctness of Bezier curve that will be reconstructed totally rely on the effectiveness our proposed network. In addition, the Hopfield neural network can be demarcated as a model of content addressable memory (CAM) [15, 16]. In lay man‟s term, we call it as Hopfield‟s brain. Hence, Hopfield‟s brain replicates our b iological brain function to store and process the memories. Besides, the learning and retriev ing data in Hopfield neural network are the fundamental aspect of content addressable memory (CAM) [11, 14]. Specifically, the Hopfield neural network is a class of recurrent autoassociative network [33]. The units in Hopfield models are predo minantly binary threshold unit [34]. Hence, the Hopfield nets will y ield a binary value such as 1 and 1. The definition for unit I‟s activation, ai are given as follows:
1 if Wij S j i ai j 1 Otherwise
(11)
where Wij is the connection strength from unit j to i .
S j is the state of unit j and i is the threshold of unit i . The connection in Hopfield net typically has no connection with itself Wii 0 and connections are symmetric or bidirectional [12, 13]. Copyright © 2016 MECS
1
2
ij
Ji
j
(12)
j
IV. BEZIERSAT IN HOPFIELD NEURAL NET WORK Neurolog ic is the blend of neural network and logic programming as a single hybrid network. Specifically, the neurologic methods can be applied in a various field such as constraint optimization problem, pattern recognition, circuit and curve reconstruction [33]. In this paper, we apply the Hopfield neural network and logic programming to reconstruct the Bezier curves according to the properties needed based on 2SAT instances. The effectiveness of our proposed paradigm (HNN2SAT) will be discussed briefly in Section VI.
13
The weight in Hopfield network is always symmet rical. The weight in Hopfield network denotes to the connection strength between the neurons. The updating rule maintains as follows:
Si t 1 sgn hi t
(13)
This property guarantee the energy will decrease monotonically even though following the activation system [14, 30]. Ho wever, it will drive the network to search for the possible minimu m energy. The following equation represents energy for Hopfield network.
E ....
1 2
W S S W S 2
ij
i
j
1
i
j
i
j
(14)
i
This energy function is vital to imp rove the degree of convergence of the proposed network [18]. Thus, the energy value is important to obtain global Bezier curves. The power o f reconstructing the curves depends on how the synaptic weights are co mputed. Hence, our proposed networks are ab le to update the weights and proceed the Bezier curves reconstruction effectively. B. Logic Programming in Hopfield Network Fundamentally, logic programming can be treated as a problem in constraint optimization outlook [22]. Hence, it can be carried out in the Hopfield neural network to obtain global solutions. This can be done by using the neurons to store the truth values of the literal and writ ing a cost function which is min imized when all clauses are satisfied [13, 29]. In other words, the core mission is to discover „models‟ corresponding to the given logic program. The important of Hopfield network in doing logic programming was brought up because of its exceptional content addressable memo ry properties during learning process. Implementation of HNN2SAT in Logic Programming. i. The 2SAT clauses are transformed into Boolean algebra. Thus, the clauses will form a formu la that will be used to check the properties of Bezier curves. In this paper, each clause represents the properties of the Bezier curve. ii. Identify a neuron to each ground neuron. Then, initialize all the possible weights to zero. I.J. Intelligent Systems and Applications, 2016, 12, 917
14
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
iii. Derive a cost function with negation of all 2SAT clauses. As an illustration, we have the following cost functions, X
1 1 1 S X and X 1 S X . Hence, 2 2
the states are represented as
S X 1 (True) and
S X 1 (False). For this scenario, the mu ltiplication signifies conjunction and addition symbolizes disjunction. iv. Co mparing the cost function with energy, E by obtaining the values of the connection strengths. (Wan Abdullah‟s Method) [9] v. Check clauses satisfaction by using exhaustiv e search. Hence, the satisfied assignments will be stored. In this paper, the satisfied assignments for the Bezier curves (correct properties) will be stored as content addressable memory (CAM). vi. The states of the neurons are randomized. The network undergoes sequences of network relaxation [14]. Calculate the resultant local field
hi t of the state. Let
say if the final state is stable for 5 runs, we consider it as final state. vii. Co mpute the corresponding final energy E of the final state via imp lementing Lypunov equation. Authenticate whether the final energy achieved is a global minimu m energy or local min ima. In Bezier curves reconstruction, the final energy depicts the correct properties of the Bezier curves. Finally, the global Bezier model and running time are co mputed for every Bezier models.
V. IMPLEMENT AT ION In this exp loration, we require methodical procedures. The procedures would work by using M icrosoft Visual C++ 2013 as a platfo rm to simu late our logic program. First of all, we generate a random program based on 2SAT clauses. In Bezier curve reconstruction, we represent each of the properties of Bezier model as the clauses. Each of the clauses form randomized 2SAT formula. Each of the clauses consist of the randomized state. The states of the 2SAT clauses (representing properties of the Bezier) will be verified and the correct clauses will be retained (train ing stage). After undergo HNN2SAT, the network reached the final states. Equation (13) is vital to ensure the network achieve stable states. Stable final states will be achieved when the state remains unchanged for 5 consecutive runs. Pinkas [18] emphasized that by permitting an ANN to evolve in time shall lead to the stable state where the energy function obtained does not change further. In this case, the corresponding final energy for the stable state will be calculated. If the difference between the final energy and the global min imu m energy is within the termination criteria, then consider the solution as global min imu m energy. 0.001 was chosen as a termination criterion since this value gave us better output accuracy. We run 100 training and 100 comb inations of neurons in order to reduce statistical error. Since the network will produce 10000 Bezier models, we calculate the percentage of Copyright © 2016 MECS
global Bezier model. Running time will be recorded fro m the start to the end of the program. In this paper, the global Bezier model is considered as the global solution.
VI. RESULT AND DISCUSSION In order to test the performance of HNN2SAT in reconstructing the Bezier curve model, we evaluated the proposed paradigm based on global Bezier model and running time. A. Global Bezier model T able 2. Global Bezier model for reconstructed linear Bezier curves Number of Reconstructed Linear Bezier Curves 20 40
Global Bezier Model (%) 100 100
60 80
98.3 96.7
100
96.2
120 140
94.9 93.7
T able 3. Global Bezier model for reconstructed quadratic Bezier curves Number of Reconstructed Quadratic Bezier Curves 20
Global Bezier Model (%) 100
40
99.9
60
98.6
80
96.6
100
95.7
120
94.8
140
92.3
T able 4. Global Bezier model for reconstructed cubic Bezier curves Number of Reconstructed Cubic Bezier Curves 20
Global Bezier Model (%) 100
40
98.9
60
98.0
80
96.1
100
95.3
120
92.5
140
90.8
Table 2, 3 and 4 illustrate the global Bezier model configuration for the d ifferent type of curve. Based on the table, we successfully reconstruct the correct Bezier model for all type curves. HNN2SAT are ab le to reconstruct 20 Bezier curves (Global Bezier model) for all order of the curve without encountering any non Bezier model (Local Bezier model). The proposed paradigm is able to train the network with the properties of the Bezier via 2SAT and successfully ret rieved the correct clause during the reconstruction phase. As the number o f reconstructed Bezier curve increased, the I.J. Intelligent Systems and Applications, 2016, 12, 917
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
complexity of the network increased. Local Bezier models are expected to emerge because the solution of HNN2SAT trapped at the suboptimal solution. This is due to spurious min ima occurred during the retrieval phase and the network recalled the wrong states for the clauses (Bezier properties) [14, 13] Consequently, the existence of local Bezier model will create a curve which does not follo w the properties of the Bezier curve. Co mparatively, cubic Bezier model produced the highest number of local Bezier curve co mpared to linear and quadratic Bezier curve. Th is is because cubic Bezier has more basis and control points compared to linear and quadratic Bezier curve. By the same token, the cubic Bezier constraints embedded in HNN2SAT will be larger and more co mp lex co mpared to linear and quadratic Bezier curve. Local Bezier curves are expected to increase if the order of the Bezier curve increased. For example, quartic Bezier curves are expected to produce more local Bezier model co mpared to cubic Bezier. In general, almost 90% of the curve (linear, quadratic and cubic) produced by HNN2SAT are global Bezier curve. B. Running Time T able 5. Running time for reconstructed linear Bezier curves Number of Reconstructed Linear Bezier Curves 20
Running T ime (s) 0.0100
40
0.2600
60
0.5238
80
1.263
100
1.640
120
2.050
140
2.503
T able 6. Running time for reconstructed quadratic Bezier curves Number of Reconstructed Quadratic Bezier Curves 20 40
Running T ime (s) 0.1500 0.2900
60
1.040
80
1.113
100
1.530
120 140
2.440 3.450
T able 7. Running time for reconstructed cubic Bezier curves Number of Reconstructed Cubic Bezier Curves 20
Running T ime (s) 0.2772
40
0.6600
60
1.318
80
1.790
100
2.175
120
3.326
140
4.100
Copyright ÂŠ 2016 MECS
15
Running time is another measure or indicator to check the effectiveness of our proposed network [15]. Table 5, 6 and 7 depict the running time for our proposed network, HNN2SAT to reconstruct different Bezier curves correctly. A closer look at the running time ind icates that the HNN2SAT has successfully reconstructed the curves within the stipulated time frame. The running time obtained in this study provides convincing evidence that the time taken for the HNN2SAT to learn and retrieve the correct curves slightly varies for the different type of Bezier curves. As a matter of fact, the training process consumed most of the running time as our proposed network will be checking the properties of the correct Bezier curves. Strict ly speaking, in order to generate the correct Bezier curves, three main properties (constraints) must be fulfilled. Hence, the verificat ion requires particular t ime to comp lete the entire process. Theoretically, as the number of Bezier curves model increases, the running time will also increase. In the previous measures, the issue under scrutiny is the existence of local Bezier in any complex model. Table 5, 6 and 7 delineates the significant differences in the running time obtained for the linear, quadratic and cubic Bezier model if the number of curves is ranging fro m 100 to 140. According to the results, the running time for linear Bezier model is faster than quadratic Bezier, especially when more curves are introduced. Conversely, the cubic Bezier model required more running time. This is due to the cubic Bezier co mp rised of more basis and control points compared to linear and quadratic Bezier curve. Thus, the training process in order to reconstruct the correct Bezier curve became much slo wer. There is overwhelming evidence corroborating the notion that more running time required for high ly comp lex Bezier model. A clear ev idence is the computation burden to reconstruct higher order curves. Hence, better optimization technique can be applied to imp rove the running time for highly co mplex Bezier models. For the time being, HNN2SAT is still the best network.
VII. CONCLUSION We have presented our proposed paradigm, namely HNN2SAT network to reconstruct various Bezier curves model. It had been presented by the computer simulat ions that our proposed model that incorporated with Hopfield neural networks were able to retrieve and reconstruct correct curves model with in the exceptional time frame. Information fro m the various Bezier models can be stored in the clausal form in Hopfield and most of the retrieved model are exact Bezier models. Hence, the proposed models are supported by the strong agreement of global Bezier model obtained and running time. This early work may have some limitations which will be addressed in future work. Finding satisfied interpretations which have been a build ing block for clause during simu lation can be very complex and detrimental. We have noted that optimized and effective searching technique should be drawn in order to find the correct clausal state. This requires further investigation. For future work, I.J. Intelligent Systems and Applications, 2016, 12, 917
16
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
approximation algorith m s uch as heuristic methods can be implemented in order to find the correct interpretation for clauses. On the separate note, we can consider variety satisfiability logic such as 3Satisfiability to represent the clauses in Hopfield network. REFERENCES [1] [2]
[3]
[4] [5] [6]
[7]
[8] [9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17] [18]
[19]
G. Farin, A history of curves and surfaces. Handbook of Computer Aided Geometric Design, Elsevier, 2002. A. Riškus, & G. Liutkus, An Improved Algorithm for the Approximation of a Cubic Bezier Curve and its Application for Approximating Quadratic Bezier Curve. Information Technology and Control, 42(4), pp. 303308, 2013. B. T. Bertka, An Introduction to Bezier Curves, BSplines, and Tensor Product Surfaces with History and Applications. University of California Santa Cruz, May 30th, 2008. R.A. Kowalski, Logic for Problem Solving. New York: Elsevier Science Publishing, 1979. T. Feder, Network flow and 2satisfiability, Algorithmica, 11(3), pp. 291319, 1994. C. H. Papadimitriou, On selecting a satisfying truth assignment. In Foundations of Computer Science, 1991. Proceedings., 32nd Annual Symposium, pp. 163169, 1991. R. Petreschi, & B. Simeone, Experimental comparison of 2satisfiability algorithms. Revue française d'automatique, d'informatique et de recherche opérationnelle, Recherche opérationnelle, 25(3), pp. 241264, 1991. W. Fernandez, Random 2SAT: Result and Problems, Theoretical Computer Science, 265, pp. 131146, 2001. W.A.T. Wan Abdullah, Logic Programming on a Neural Network. M alaysian Journal of computer Science, 9 (1), pp. 15, 1993. J. J. Hopfield, D. W. Tank, Neural computation of decisions in optimization problem, Biological Cybernatics, 52, pp. 141152, 1985. S. Haykin, Neural Networks: A Comprehensive Foundation, New York: M acmillan College Publishing, 1999. G. Pinkas, R. Dechter, Improving energy connectionist energy minimization, Journal of Artificial Intelligence Research, 3, pp. 22315, 1995. S. Sathasivam, Upgrading Logic Programming in Hopfield Network, Sains M alaysiana, 39, pp. 115118, 2010. S. Sathasivam, Energy Relaxation for Hopfield Network with the New Learning Rule, International Conference on Power Control and Optimization, pp. 15, 2009. S. Sathasivam, P.F. Ng, N. Hamadneh, Developing agent based modelling for reverse analysis method, 6 (22), pp. 42814288, 2013. S. Sathasivam, Learnin g in the Recurrent Hopfield Network, Proceedings of the Fifth International Conference on Computer Graphics, Imaging and Visualisation, pp. 323328, 2008. R. Rojas, Neural Networks: A Systematic Introduction. Berlin: Springer, 1996. G. Pinkas, Propositional nonmonotonic reasoning and inconsistency in symmetric neural networks. Washington University, Department of Computer Science, 1991. K. G. Jolly and R. S. Kumar, A Bezier curve based path planning in a multiagent robot soccer system without violating the acceleration limits, Robotics and Autonomous System, 57(1), pp. 2333, 2009.
Copyright © 2016 MECS
[20] G. Farin, Curves and Surfaces for ComputerAided Geometric Design: A Practical Guide, Elsevier, 2014. [21] Patterson and R. Richard, Projective Transformations of the Parameter of a BernsteinBezier Curve, ACM Transactions on Graphics, TOG, 4(4), pp. 276290, 1985. [22] F. A. Robin, Interactive Interpolation and Approximation by Bezier Polynomials, 15(1), pp. 7179, 1972. [23] Bohm, Wolfgang, G. Farin, J. Kahmann, A survey of curve and surface methods in CAGD: Computer Aided Geometric Design 1, 1, pp. 160, 1984. [24] A. M arc, Linear Combination of Transformation, ACM Transaction on Graphics, TOG 21(3), pp. 380387, 2002. [25] S.M . Hu, G. Z. Wang, T.G. Jin, Properties of two types of generalized Ball curves: ComputerAided Design 28, 2, pp. 125133, 1996. [26] C.H. Chu, H. S. Carlo, Developable Bezier patches: properties and design: ComputerAided Design 34, 7, pp. 511527, 2002. [27] J. Gu, Local Search for Satisfiability (SAT) Problem, IEEE Transactions on Systems, M an and Cybernetics, vol. 23 pp. 11081129, 1993. [28] R. Puff, J. Gu, A BDD SAT solver for satisfiability testing: An industrial case study , Annals of M athematics and Artificial Intelligence, 17 (2), pp. 315337, 1996. [29] A. Cimatti, M . Roveri, Bertoli. P, Conformant planning via symbolic model checking and heuristic search, Artificial Intelligence, 159 (1), pp. 127206, 2004. [30] A. D. Pwasong and S. Sathasivam, Forecasting Performance of Random Walk with Drift and Feed Forward Neural Network M odels, International Journal of Intelligent System and Application, vol. 23 pp. 4956, 2015. [31] D. Vilhelm, J. Peter, & W. M agnus, Counting models for 2SAT and 3SAT formulae. Theoretical Computer Science, 332 (1), pp. 265291, 2005. [32] B. Sebastian, H. Pascal and H. Steffen, Connectionist model generation: A firstorder approach, Neurocomputing 71(13) (2008) 24202432. [33] H. Asgari, Y. S. Kavian, and A. M ahani, A Systolic Architecture for Hopfield Neural Networks. Procedia Technology, 17 (2014) 736741. [34] Y. L. Xou and T. C. Bao, Novel global stability criteria for highorder Hopfieldtype neural networks with timevarying delays, Journal of M athematical Analysis and Applications 330 (2007) 14415. [35] K. Iwama, “CNF Satisfiability Test by Counting and Polynomial Average Time,” SIAM , Journal of Computer, vol. 4, no. 1, pp. 385391, 1989.
Authors’ Profiles Mohd S hareduwan bin M. Kasihmuddin received his M Sc (2014) and BSc(Ed) (2013) from Universiti Sains M alaysia. He is currently pursuing Ph.D degree with School of M athematical Science, Universiti Sains M alaysia Penang M alaysia. His current research interests include neuroheuristic method, constrained optimization, neural network and logic programming.
I.J. Intelligent Systems and Applications, 2016, 12, 917
Bezier Curves Satisfiability M odel in Enhanced Hopfield Network
17
Mohd Asyraf Mansor was born in Sarawak, M alaysia in 1990. He obtained his M Sc (2014) and BSc(Ed) (2013) from Universiti Sains M alaysia. He is currently pursuing Ph.D degree at School of M athematical Science, Universiti Sains M alaysia. His current research interests include evolutionary algorithm, satisfiability problem, neural networks, logic programming and heuristic method especially Artificial Immune System.
S aratha S athasivam is a lecturer in the School of M athematical Sciences, Universiti Sains M alaysia. She received her M Sc and BSc(Ed) from Universiti Sains M alaysia. She received her Ph.D at Universiti M alaya, M alaysia. Her current research interest are neural networks, agent based modeling and constrained optimization problem.
How to cite this paper: M ohd Shareduwan M . Kasihmuddin, M ohd Asyraf M ansor, Saratha Sathasivam, "Bezier Curves Satisfiability M odel in Enhanced Hopfield Network", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.917, 2016. DOI: 10.5815/ijisa.2016.12.02
Copyright ÂŠ 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 917
I.J. Intelligent Systems and Applications, 2016, 12, 1825 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.03
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks Vu Chien Thang Faculty of Electronics and Communications Technology , University of Information and Communications Technology, Thai Nguyen, 250000, Viet Nam Email: vcthang@ictu.edu.vn
Nguyen Van Tao Faculty of Information Technology, University of Information and Communications Technology, Thai Nguyen, 250000, Viet Nam Email: nvtao@ictu.edu.vn
Abstract—In the near future, IPbased wireless sensor networks will play a key ro le in several application scenarios such as smart grid, smart ho me, healthcare, and building automat ion... An IPv6 routing protocol is expected to provide internet connectivity to any IPbased sensor node. In this paper, we propose IRPL protocol for IPbased wireless sensor networks. IRPL protocol uses a combination of two routing metrics that are the link quality and the remain ing energy state of the preferred parent to select the optimal path. In IRPL protocol, we combine t wo metrics based on an alpha weight. IRPL protocol is implemented in ContikiOS and evaluated by using simu lation and testbed experiments . The results show that IRPL protocol has achieved better network lifetime, data delivery rat io and energy balance co mpared to the traditional solution of RPL protocol. Index Terms—Imp roved IPv6 routing protocol, wireless sensor networks, contiki operating system, network performance evaluation.
I. INT RODUCT ION Internet of Things (IoT) [1] is a scenario in which the billions of devices are connected together and each device has a unique global address. In this scenario, IPv6 Wireless Sensor Networks (WSNs) have a key ro le since they will be used to collect several environ ment informat ion. The applications of W SNs such as healthcare, build ing automation, and smart grid,...[2] with mu ltihop commun ication model that use IEEE 802.15.4 standard will also be a part of IoT [3]. At the beginning, the wireless sensor network community rejected the IP architecture based on the assumption that it would not meet the challenges of wireless sensor networks. In fact, many have moved to IP because of the interoperability with existing systems and the wellengineered architecture based on the endtoend architecture. Therefore, the Internet Engineering Task Force (IETF) has created 6LoWPAN (IPv 6 over LowPower W ireless
Copyright © 2016 MECS
Personal Area Net works) and Ro LL (Routing over Lo wpower and Lossy networks) working groups to standardize the IP arch itecture for WSNs [4]. While the work fro m the 6LoWPA N working group opened the possibility of using IPv 6 in IEEE 802.15.4 networks, standardizing a routing protocol was outside the scope of that working group. This led to the creation of the Ro LL working group in 2008. Ro LL Working Group has studied on standardization issues of IPv6 routing protocol for constrained devices over lowpower and lossy networks. This group has proposed RPL protocol as an IPv6 routing protocol for WSNs. In [5], Nguyen Thanh Long et al. p resented a comprehensive study of the performance of RPL when compared to collect ion tree protocol. The study shows that the number of switching parents (Churn) in the RPL network is very low. Currently, RPL uses ETX (expected transmission) as its routing metric to avoid lossy lin ks [6]. However, ETX doesn’t address the problem of energy balancing. Therefo re, RPL is prone to the hot spot problem: certain nodes belong the routes that have good lin k quality will carry much heavier trans mission load than other nodes. These nodes are likely to run out of battery faster than the ordinary nodes, which may create holes and undermine the network lifetime. In this paper, we will propose improved RPL (IRPL) protocol in order to overcome this disadvantage of RPL protocol. IRPL uses two routing metrics that are ETX and the energy state of sensor nodes to choose the optimal path. The contribution of this paper is threefold : Firstly, we propose a combined routing metric and show how to estimate it for each node; Secondly, we present an algorithm to choose the optimal path based on this routing metric; Finally, we imp lement and evaluate our proposal by using simulation and testbed experiments. The remainder of this paper is organized as follows. Section II shows the related works. Section III introduces IRPL’s imp lementation principle. In section IV, the performance of IRPL protocol is evaluated and co mpared to the original RPL protocol; Finally, we conclude the paper.
I.J. Intelligent Systems and Applications, 2016, 12, 1825
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
II. RELAT ED W ORKS Routing in WSNs has been extensively studied in the last decade. Since most of the sensor nodes are battery powered, then a good routing protocol should save energy. In this section, we will present the descriptions and the implementations of RPL. Then, we will present the related works specifically for energy efficiency at the routing layer in WSNs. A. Descriptions of RPL RPL is designed for WSNs where sensor nodes are interconnected by wireless and lossy links. The lossy nature of these links is not the only WSN characteristic that drove the design decisions of RPL. Because resources are scarce, the control traffic must be as tightly bounded as possible. In these networks , the data traffic is usually limited and the control traffic should be reduced whenever possible to save bandwidth and energy. RPL is a distance vector protocol that builds and uses DODA G in the network to perform routing [4]. In which, a DODA G is a network topology where all lin ks between nodes in the DODA G has specified direct ion, toward the DODAG root. Fig. 1 illustrates an RPL DODAG.
Fig.1. An RPL DODAG.
First, one or more nodes are configured as DODA G roots by the network ad ministrator. RPL defines three new ICMPv 6 messages called DODA G Informat ion Object (DIO), DODA G Information Solicitation (DIS), and Destination Advertisement Ob ject (DAO) messages . DIO messages are sent by nodes to advertise informat ion about the DODA G, such as the DODA GID, the OF, DODA G rank, along with other DODA G parameters such as a set of path metrics and constraints. DIS messages are used to solicit DIO fro m an RPL node. A node may use DIS messages to probe its neighborhood for nearby DODAGs. RPL uses “up” and “down” directions terminology. The up direction is fro m a leaf toward the DODA G root, whereas down refers to the opposite direction. The usual terminology of parents/children is used. RPL also introduces the “sibling”: two nodes are siblings if they have the same rank in the DODA G. The parent of a node in the DODA G is the immediate successor within the Copyright © 2016 MECS
19
DODA G in the up direction, whereas a DODA G sibling refers to a node at the same rank. B. Implementations of RPL Currently, several uIP stacks are available such as uIPv6 in ContikiOS; Blip in TinyOS [7]... These uIP stacks are lightweight and can be ported in several microcontrollers of sensor nodes including the MSP430 fro m Texas Instruments and the AVR ATMega128 fro m Atmel. Therefore, all the wireless sensor networks can be connected to the Internet as any computer devices. The first draft of RPL was launched in August 2009 by IETF. One year later, the implementations of RPL were performed fo r TinyOS and ContikiOS. In [8], J. Tripathi et al. presented a performance evaluation of RPL. The authors use OMNET++ to simulate RPL. So me metrics for evaluation of RPL also are p resented such as routing table size, expected transmission count, control packet overhead, and loss of connectivity. At the same time, Nicolas Tsiftes et al. presented the ﬁrst experimental results of RPL which they obtained with their ContikiRPL imp lementation [9]. They evaluate the power efﬁciency of ContikiRPL by running it in a 41node simu lation and in a smallscale 13node Tmote Sky deployment in an ofﬁce environment. Their results show several years of network lifet ime with IPv6 routing on Tmote Sky motes. In [10], the authors presented a framework for simulat ion, experimentation, and evaluation of routing mechanisms for low power IPv6 networking in Contiki. Th is framework provides a detailed simulat ion environ ment for lowpower routing mechanis ms and allo ws the system to be directly uploaded to a physical testbed for experimental measurements. In [7], JeongGil Ko et al. presented two interoperable implementations of RPL for TinyOS and ContikiOS. They demonstrate interoperability between two implementations. Currently, RPL uses ETX as its routing metric to avoid lossy links. ETX is the inverse of the product of the forward delivery rat io, D f and the backward delivery ratio, Db , which takes into account link asymmetry.
ETX
1 D f .Db
(1)
Db refers to the packet reception ratio, while D f refers to the acknowledg ment reception ratio. Routing protocols based on the ETX met ric provide highthroughput routes on mult ihop wireless networks [11], since ETX minimizes the expected total number of packet transmissions required to successfully deliver a packet to the destination. This will result in a min imu m energy path. However, by using the min imu m energy path to route all the packets, the nodes on that path will quickly run out of energy. It will not improve the lifetime of the whole network and the resulted topology will not be energy balanced. In order to avoid this problem, we propose IRPL. This is done by having residual energy as a decision factor in the routing tables and this informat ion is exchanged between the neighboring nodes. I.J. Intelligent Systems and Applications, 2016, 12, 1825
20
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
C. Energy Aware Routing Many energyaware routing protocols have been proposed to min imize the energy consumption and to prolong the network lifetime. In [12], Kamgueu et al. proposed to use RPL with a residual energy metric. However, they do not consider the rad io link quality. Oana Iova et al. [13] proposed the Expected LifeTime (ELT) routing metric to estimate the lifetime of bottlenecks. The authors take into account both the amount of traffic and the link reliability to estimate how much energy such a bottleneck consumes on average. RPL was used as the routing protocol in [14] where the forwarding load is weighted between the members of the parent list. The weighting is based on the members’ residual energy. The transmission range dynamically adjusted to maintain k parents. In [15], Rahul C. Shah et al. proposed Energy Aware Routing (EA R). EAR also maintains a set of good paths instead of a single optimal path. It introduces an energy metric wh ich is determined by both the cost to deliver a packet and the residual power of the intermed iate node. EAR requires a node to choose one path from those paths probabilistically. The probability assigned to each path is determined by th e energy metric of the path. IRPL is different fro m EA R in several aspects. First, EA R’s routing decision is based on the residual energy of paths while IRPL’s is based on the residual energy indicator of indiv idual nodes. Fro m our perspective, EAR ignores the energy differences of sensor nodes on the same path. A path is abundant in residual energy does not mean all the nodes on the path is abundant in residual energy. Second, EA R’s energy cost metric is based on the assumption that the accurate residual energy and transmission cost of energy can be obtained by sensor nodes. However, this is not true for some sensor platforms. In IRPL, the residual energy is estimated by software and this mechanis m can be easily added to existing sensor platforms. The detail of IRPL will be presented in section III.
run out of energy. B. Design Solution The residual energy of a sensor node is defined by (2). In (2), Eresidual , E0 , Econsumption are respectively the residual energy, the initial energy and the energy consumption of the sensor node.
Eresidual E0 Econsumption
(2)
The total energy consumption of the sensor node is defined as [16]:
Econsumption U ( I ata Il tl It tt I r tr I citci ) (3) i
Where U is the supply voltage, Ia is the consumption current of the microcontroller when running, ta is the time in wh ich the microcontroller has been running, Il and tl are the consumption current and the time of microcontroller in low power mode, It and t t are the consumption current and the time of the co mmun ication device in transmit mode, Ir and tr are the consumption current and the time of co mmun ication device in receive mode, Ici and t ci are the consumption and the time of other components such as sensors and LEDs... In this paper, we evaluate the performance of IRPL on TUmote (Thai Nguyen University mote). TUmote is a hardware p latform for extremely lo w power, h igh datarate sensor network applications. Fig. 2 shows the structure of TUmote.
CC2420
MSP430 F1611
SHT11
Battery
Fig.2. Structure of T Umote.
III. IRPL PROT OCOL In this section, we will present the design goal and some challenges of IRPL design. Then, we will present our solution for implementation of IRPL. A. Design Goal and Challenges The main goal of IRPL design is to balance the energy of sensor nodes that belong to the routes having good link quality and increase the lifet ime of sensor nodes. Some challenges of IRPL design are: First, we need to determine the residual energy indicator of each sensor node. The mechanism to determine the energy ind icator must be easily added to existing hardware and software designs, without any additional hardware cost. Second, we need to propose an algorith m to choose the optimal path based on two routing metrics that are ETX and EI of forwarded nodes. The chosen route must have good link quality and avoid choosing sensor nodes that Copyright © 2016 MECS
Table 1 shows our energy model, where the consumption currents are fro m chip manufacturer data sheets. In the energy model of TUmote, we on ly consider on the main energy consumptions that are the radio transceiver, the microcontroller, and the sensor, we ignore other small energy consumptions. T able 1. Energy Model of T Umote Component
State
Current
Active
1,95 mA
Low power mode
0,0026 mA
T ransmit (0 dBm)
17,4 mA
Listen
19.7 mA
Active
0,55 mA
MSP430 F1611
CC2420 SHT 11
I.J. Intelligent Systems and Applications, 2016, 12, 1825
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
21
The energy indicator of the sensor node is defined by this equation:
EI (%)
Eresidual .100% E0
(4)
Fig. 3 shows the imp lementation of IRPL in Contiki operating system [17, 18]. We expand the routing table in ContikiRPL to store the informat ion of residual energy indicator of neighbors. ContikiRPL uses ETX and EI metrics to choose the optimal path. The link estimator module will estimate the lin k quality. The energy estimator module will determine the residual energy indicator of the sensor node. Application Data Packets
IV. PERFORMANCE EVALUAT ION OF IRPL
UDP/TCP Data Packets
A. Evaluation Metrics
ICMPv6 (DIO, DIS, DAO) Neighbor Discovery uIPv6
ContikiRPL
Route Install Link Estimate
Data and Control Packets MAC, IEEE 802.15.4 PHY
Energy Estimate
Link Estimator
Energy Estimator
Link Result
We evaluate the performance of IRPL through a set of metrics that can outline the most significant features of the protocol. Data delivery ratio: The first metric is the data delivery ratio (DDR). We define DDR as the ratio between the number of data packets received at the DODA G root and the number of data packets sent by nodes in the network.
Fig.3. Implementation of IRPL in ContikiOS.
In this paper, we propose a solution for co mb ination of ETX and EI routing metrics by this equation:
metricETX _ EI (%)
ETX 100 (1 )(100 EI ) ETX max
(5)
Where α is the weight that allows adjusting between ETX and EI metrics in order to calculate the co mbined routing metric. The value of α weight ranges fro m 0 to 1; ETXmax is the maximu m value of route quality in the network. The co mb ined routing metric is carried by the reserved field in DIO message. Fig. 4 shows the structure of DIO message [19].
DDR(%)
N received .100% N data
In (6), Nreceived is the number of data packets received at the DODA G root; Ndata is the number of data packets sent by all nodes in the network. Clearly, DDR equals to 1 indicates that the network can deliver all the data to the DODAG root. Energy balance of routing protocol: We evaluate the energy balance of routing protocol based on the residual energy indicator. We calculate the energy balance indicator (EBI) as in (7). EBI
N
( EI EI )
The goal of IRPL is to find the optimal path based on minimizing the combined routing metric. Algorith m 1 shows the pseudocode to select the preferred parent.
Copyright © 2016 MECS
2
(7)
i
i 1
Fig.4. Structure of DIO message.
(6)
In (7), EI reflects the average energy indicator of the whole sensor nodes. Network lifetime: The main objective of any routing protocol is to extend the lifet ime of WSNs. The network lifetime can be defined as the interval of t ime, starting with the first transmission in the wireless sensor networks and ending when the alive nodes ratio (ANR) falls below a specific threshold, which is set according to the type of application (it can be either 100% or less) [20].
ANR (%)
N alive _ nodes N
.100%
(8)
I.J. Intelligent Systems and Applications, 2016, 12, 1825
22
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
In (8), Nalive_nodes is the number of alive nodes in the network; N is the total nu mber o f nodes in the network. If the threshold of ANR is set to 100%, then once the first node expires the network is considered expired as well.
Fig. 6, 7, 8 show respectively the comparison between IRPL and RPL based on the alive node ratio, the data delivery ratio, and the energy balance indicator with different α.
B. Simulation and Evaluation This section presents the performance evaluation of IRPL. We used the COOJA network simulator [18, 21] to simu late IRPL and analyze the results. In order to evaluate the benefits of IRPL for WSNs, we co mpared its performance to RPL’s. We evaluated the performance of two protocols with the same simulation scenario. Fig. 5 shows the network topology of 26 nodes which are placed randomly in the sensor field and the node 1 is the DODA G root (Sin k). The size o f the network is 100m x 100m. Each node generates data packets for every 15 seconds. The sink collects the data packets and forwards the data packets to the PC. Table 2 shows the simulation scenario in detail. We changed the alpha weight to evaluate the in fluence of this weight on the performance of the network. The va lue of α was chosen in the range of 0.8, 0.85, 0.9, 0.95.
Fig.6. Comparison of alive node ratio in the simulation.
Fig.7. Comparison of data delivery ratio in the simulation.
Fig.8. Comparison of energy balance indicator in the simulation. Fig.5. Network topology in the simulation.
T able 2. Simulation Scenario Parameter Radio model Number of nodes
Value UDI (Unit Disk Graph with Distance Interference) [22] 26
Network size
100m x 100m
Communication range of node
T ransmission range: 30m, Interference range: 50m
Initial energy
10J
Data packet interval
15s
Data packet initialization
All nodes except the DODAG root
MAC protocol
CSMA/ContikiMAC [23]
Copyright © 2016 MECS
The simulation results show that IRPL with the α weight of 0.9 provides the best performance in terms of the alive node ratio (in Fig. 6). In this paper, we choose the threshold of ANR is 100% to determine the network lifetime. With this threshold, the network lifetime increases by up to 46% co mpared to the original RPL protocol. The simu lation results in Fig. 7 also show that IRPL with the α weight of 0.9 achieves higher data delivery rat io than RPL does. Fro m Fig. 6 and 7, we can see that IRPL with the α weight of 0.9 guarantees the best balance between the network lifetime and the data delivery ratio. C. Testbed Experiments and Evaluation IRPL is evaluated through a testbed (smallscale). The testbed results are used to calibrate the simu lation results, I.J. Intelligent Systems and Applications, 2016, 12, 1825
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
and also to analyze the performance of IRPL in real experiments. The experiments were carried out in the 1st floor of a smart ho me. The experiments consisted of 9 TUmotes in a grid topology and the sink on the edge, as depicted in Fig. 9. The experiments were carried out in an indoor environment. In this scenario, the nodes were deployed in rooms of the smart ho me. The experiments were conducted in a real world environment in the presence of continuous movement of people and other wireless network devices, which caused noise and interference. Testbed experiments were also conducted to evaluate IRPL and co mpare its performance to RPL’s in terms of the network lifetime, the data delivery rat io, and the energy balance indicator in experiments into the real environment. The experiments were set up to allow each node to send a data packet to the sink with an interval of 30 seconds. In Fig. 9, the sink collects the data packets and forwards the data packets to the PC. Fro m the simu lation results, we chose IRPL with the α weight of 0.9. Table 3 shows the testbed scenario.
23
energy consumption for all sens or nodes.
Fig.10. Comparison of alive node ratio in the testbed.
Fig.11. Comparison of data delivery ratio in the testbed.
Fig.9. Network topology in the testbed. Fig.12. Comparison of energy balance indicator in the testbed. T able 3. T estbed Scenario Parameter
Value
T ransmission environment
Indoor
Number of nodes
9
T ransmit power
0 dBm
Initial energy
10J
Data packet interval
30s
Data packet initialization
All nodes except the sink
MAC protocol
CSMA/ContikiMAC [23]
Fig. 10, 11, 12 show respectively the comparison between IRPL and RPL for this testbed. The testbed results are also similar to the simulation results. Fig. 10 shows that IRPL increases the network lifetime by up to 17% co mpared to RPL. This is due to the fact that IRPL uses a load balancing scheme that provides uniform Copyright © 2016 MECS
Fig. 11 shows the results of the data delivery ratio when using RPL o r IRPL as the routing protocol. It is important to point out that IRPL achieves higher data delivery ratio than RPL does. This is due to the fact that the alive node ratio is higher when using IRPL. Thus, the more nodes survive in the network, the mo re data packets received at the sink. Fig. 12 shows the energy balance indicator for this testbed. RPL presents unbalanced energy consumption between sensor nodes. IRPL does not increase energy consumption while imp roving the energy balance indicator.
V. CONCLUSIONS In this paper, we have proposed IRPL. IRPL uses two routing metrics that are the link quality and the residual energy indicator to select the optimal path. IRPL was evaluated by using simulation and testbed experiments to I.J. Intelligent Systems and Applications, 2016, 12, 1825
24
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
show its effects and benefits when co mpared to existing solution. Testbed results were used to calibrate and confirm the accuracy o f the simulation experiments, as well as present the impact of IRPL in the real scenario. At the same time, the simulation experiments were useful to evaluate IRPL and compare it to RPL in terms of the alive node ratio, the data delivery ratio, and the energy balance indicator in the largescale scenario. In the largescale network, the simu lation results showed that IRPL increases the network lifetime by up to 46% compared to RPL. REFERENCES Luigi Atzori, Antonio Iera, Giacomo M orabito, “The Internet of Things: A survey,” Computer Networks, 54, pp. 27872805, June 2010. [2] Hamida Qumber Ali, Sayeed Ghani, “A Comparative Analysis of Protocols for Integrating IP and Wireless Sensor Networks,” Journal of networks, Vol 11, No.1, January 2016. [3] M d. Sakhawat Hossen, A. F. M . Sultanul Kabir, Razib Hayat Khan and Abdullah Azfar, “Interconnection between 802.15.4 Devices and IPv6: Implications and Existing Approaches,” IJCSI Int. Journal of Computer Science Issues, Vol. 7, Issue 1, January 2010. [4] JeongGil Ko, Andreas Terzis, Stephen DawsonHaggerty, David E. Culler, Jonathan W. Hui, Philip Levis, “Connecting LowPower and Lossy Networks to the Internet,” IEEE Communications Magazine, pp. 96 – 101, April 2011. [5] Nguyen Thanh Long, Niccolò De Caro, Walter Colitti, Abdellah Touhafi, Kris Steenhaut, “Comparative Performance Study of RPL in Wireless Sensor Networks,” in Proceedings of 19th IEEE Symposium on Communications and Vehicular Technology in the Benelux, 2012. [6] JeongGil Ko, Joakim Eriksson, Nicolas Tsiftes, Stephen DawsonHaggerty, JeanPhilippe Vasseur, M athilde Durvy, Andreas Terzis, Adam Dunkels, and David Culler, “Beyond Interoperability  Pushing the Performance of Sensor Network IP Stacks,” in Proceedings of the 9th ACM Conference on Embedded Networked Sensor Systems (sensys'11), November 2011. [7] JeongGil Ko, Joakim Eriksson, Nicolas Tsiftes, Stephen DawsonHaggerty Andreas Terzis, Adam Dunkels and David Culler, “ContikiRPL and TinyRPL: Happy Together,” in Proceedings of the workshop on Extending the Internet to Lowpower and Lossy Networks (IP+SN 2011), Chicago, IL, USA, April 2011. [8] J. Tripathi, J. C. de Oliveira, J. P. Vasseur, “A Performance evaluation study of RPL: Routing protocol for low power and lossy networks”, in Proceedings of the 44th Annual Conference on Information Sciences and Systems, M arch 2010. [9] N. Tsiftes, J. Eriksson, and A. Dunkels, “LowPower Wireless IPv6 Routing with ContikiRPL,” in Proceedings of the International Conference on Information Processing in Sensor Networks (ACM/IEEE IPSN), Stockholm, Sweden, April 2010. [10] N. Tsiftes, J. Eriksson, N. Finne, F. Österlind, J. Höglund, and A. Dunkels, “A Framework for LowPower IPv6 Routing Simulation, Experimentation, and Evaluation,” in Proceedings of the conference on Applications, technologies, a rchitectures, and protocols for computer communications (ACM SIGCOMM), New Delhi, India, [1]
Copyright © 2016 MECS
August 2010. [11] De Couto D, Aguayo D, Bicket J, Morris R , “A highthroughput path metric for multihop wireless routing,” in Proceedings of the 9th Annual International Conference on Mobile Computing and Networking, New York, 2003. [12] P. Kamgueu, E. Nataf, T. Ndié, O. Festor, “Energy Based Routing M etric for RPL,” Research Report RR8208, INRIA, 2013. [13] O. Iova, Fabrice Theoleyre, and Thomas Noel, “Improving network lifetime with energy balancing routing: Application to RPL,” in Proceedings of Wireless and Mobile Networking Conference, 2014. [14] M . N. Moghadam, H. Taheri and M . Karrari, “M inimum cost load balanced multipath routing protocol for low power and lossy networks,” Wireless Networks, Volume 20, Issue 8, pp. 24692479, 2014. [15] R. C. Shah and J. M . Rabaey, “Energy aware routing for low energy adhoc sensor networks,” in Proceedings of Wireless Communications and Networking Conference, 2002. [16] Adam Dunkels, Fredrik Osterlind, Nicolas Tsiftes, Zhitao He, “Softwarebased Online Energy Estimation for Sensor Nodes,” in Proceedings of the 4th workshop on Embedded networked sensors, 2007. [17] A. Dunkels, B. Grönvall, and T. Voigt, “Contiki – a lightweight and flexible operating system for tiny networked sensors,” in Proceedings of EmNets, pp. 455462, 2004. [18] J. J. P. C. Rodrigues and P. A. C. S. Neves, “A survey on IPbased wireless sensor network solutions,” Int. Journal of Communication Systems, Int. J. Comm Syst. 2010. [19] Tsvetko Tsvetkov, “RPL: IPv6 Routing Protocol for Low Power and Lossy Networks,” Seminar SN SS2011 : Network Architectures and Services, pp. 59–66, July 2011. [20] Roberto Verdone, Davide Dardari, Gianluca M azzini, Andrea Conti, “Wireless Sensor and Actuator Networks: Technologies, Analysis and Design,” Academic Press, ISBN10: 0123725399, 2008. [21] Fredrik Österlind, Adam Dunkels, Joakim Eriksson, Niclas Finne, and Thiemo Voigt, “Crosslevel sensor network simulation with cooja,” in Proceedings of the First IEEE International Workshop on Practical Issues in Building Sensor Network Applications (SenseApp 2006), Tampa, Florida, USA, pp. 641648, November 2006. [22] Azzedine Boukerche, “Algorithms and Protocols for Wireless Sensor Networks,” John Wiley & Sons Inc., ISBN: 9780470396360, 2008. [23] A. Dunkels, “The ContikiM AC Radio Duty Cycling Protocol,” SICS technical report, December 2011.
Authors’ Profiles Vu Chien Thang received the M Sc degree in Electronics and Communication Technology in 2009 from Hanoi University of Science and Technology and PhD in Telecommunication Engineering in 2015 from Vietnam Research Institute of Electronics, Informatics and Automation. He is currently a lecturer at Thainguyen University of Information and Communication Technology. His research interests include wireless sensor networks, internet of things, embedded systems.
I.J. Intelligent Systems and Applications, 2016, 12, 1825
A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks
25
Nguyen Van Tao received the PhD degree in Information Technology in 2009 from Institute of Information Technology, Vietnam. He is currently a lecturer at Thainguyen University of Information and Communication Technology. His research interests include internet of things, image processing.
How to cite this paper: Vu Chien Thang, Nguyen Van Tao, "A Performance Evaluation of Improved IPv6 Routing Protocol for Wireless Sensor Networks", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.1825, 2016. DOI: 10.5815/ijisa.2016.12.03
Copyright ÂŠ 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 1825
I.J. Intelligent Systems and Applications, 2016, 12, 2636 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.04
A Review of Methods of Instancebased Automatic Image Annotation Morad Derakhshan Graduate student of Software, Department of Engineering, Sanandaj Branch, Islamic Azad University, Sanandaj, Iran Email: morad.derakhshan@gmail.com
Vafa Maihami Faculty member, Department of Engineering, Sanandaj Branch, Islamic Azad University, Sanandaj, Iran Email: maihami@iausdj.ac.ir
Abstract—Today, to use automatic image annotation in order to fill the semantic gap between low level features of images and understanding their information in retriev ing process has become popular. Since automatic image annotation is crucial in understanding digital images several methods have been proposed to automatically annotate an image. One of the most important of these methods is instancebased image annotation. As these methods are vastly used in this paper, the most impo rtant instancebased image annotation methods are analy zed. First of all the main parts of instancebased automatic image annotation are analyzed. Afterwards, the main methods of instancebased automatic image annotation are reviewed and co mpared based on various features. In the end the most important challenges and openended fields in instancebased image annotation are analyzed. Index Terms—Automat ic Image Annotation, InstanceBased Nearest Neighbor, Semantic Gap, Vot ing Algorithm.
process is one of the applications of Machine Vision in image retrieving systems and it is used to organize and locate the existing images in sets. In textbased methods the retrieving p rocess is based on texts and keywords written for each image. In this method whenever a query is received fro m a user the images enjoying that kind of query are retrieved. Here in this paper the main parts of instancebased image automat ic annotation are first analyzed.Afterwards, the main methods of instancebased automatic image annotation are reviewed and compared based on different features. The main existing challenges in this field are recognized and analyzed. The rest of the paper is as fo llo ws: in the second part the main parts of instancebased automatic image annotation are briefly carried out, the most important and wellknown algorithms of instancebased automatic image annotation are reviewed and compared with each other. And in the end the conclusion and open ended fields are proposed.
I. INT RODUCT ION
II. A REVIEW OF T HE M AIN PART S OF INST ANCE BASED IMAGE A NNOT AT ION
Today, due to the increasing growth of d igital images and the need to manage and retrieve them image annotation has become a dynamic field in research. The aim o f annotation is to accompany the words denoting the mean ings and concepts with the image. Interpret ing this volume of images by human being is impossible, costly, and time consuming, so to automate the annotation process seems to be essential. However, information and features extracted fro m the images do not always reflect the Image content and the semantic gap as “the lack of coincidence between the information that one can extract fro m the visual data and the interpretation that the same data have for a user in a given situation” is known as the main challenge of automatic systems. Recently, researches have focused on Semisupervised systems so as to fill the semantic gap by helping data p rod uced by users . Too man y met ho ds h av e b een p ro posed in th is field. Auto mat ic Image annot at io n
The main parts of instancebased image annotation are shown in figure 1. In these systems the images of the data set are first read offline and the intended features of the images are extracted and a database containing feature Vectors are saved. In the next phase the image which is intended to be annotated is received fro m the input online as the query image. Again and identical to the offline phase, the intended features are extracted and there will be a vector of feature. In order to obtain intended tags fro m the existing images in the dataset the feature vector of the query image is co mpared with feature vectors of the images of the set for being similar by the help of Similarity measures to find the nearest image by the help of the nearest neighbor method. In the next phase the best tags are obtained for the query image by using methods such as voting the tags of the near images. Next, each part of instancebased automatic image annotation is analy zed separately.
Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 2636
A Review of M ethods of Instancebased Automatic Image Annotation
Offline
Online
Dataset
Get the query image from user
Features extraction Extract image features
Create database of features
Create a query feature vector image
Similarity measure
Obtaining nearest neighbor
Apply a method of automatic annotation (voting neighbor, tag rating, …) neighbor ratings , …) Assign optimization tags to the image query
Display output Fig.1. Schematics of annotations system
A. The images database In instancebased automatic image annotation systems the images database plays a crucial role in precise automatic annotation. There various databases in this field that they each have different images and tags. Following this section the most important set of images which is used in various studies is analyzed:
NUSWIDE: This set has been prepared by Singapore National University and contains 269648 images. Of course this set is a set of images features vector with 501 d imensions. This set has become a reference set of images annotation. This set contains 81 tags. Mir Flickr: Mir Flickr database produced in Leiden University contains 25000 images and 38 tags and it is especially allocated for image retriev ing enhancement. The images in Flickr have collected as metadata EXIF and they are readily available in text files. MIRFLICKR25000: A big effort has been made to make an image set and evolving ideas. Image set prepare metadata and annotating. If one inserts one's email address before downloading, he/she can receive the latest updates. Corel5k: It contains 5000 images and 374 tags. However, Corel10k contains 10000 images in 100
Copyright © 2016 MECS
27
groups. Every category contains 100 images of size 192×128 or 128×192 in JPEG fo rmat images. This set is only used for scientific co mmunication not in commercial properties. IAPR TC12: This set contains 20000 images and 291 tags. TC12 is used for evaluating image automatic annotation methods and studying their effects on mult imedia informat ion retrieving. The images are segmented and features are ext racted fro m each segment and every single segment is tagged. Annotation is carried out in the region according to annotation hierarchy and spatial relationships information. Each image is manually segmented and the resultant regions have been annotated according to predefined words of tags. Wang: Exists 1000 color images in this data set which are organized in 10 groups. Each group contains images and textual description for a category of butterflies collected fro m Google through querying with their scientific names of the species, for instance "Danaus plexppus". They are also manually filtered for those depicting the butterfly of interest. The textual descriptions were obtained from the eNature online nature guide for every single butterfly. ImageNet: Th is set has been organized according to WordNet hierarchy. Each meaningful concept in WordNet, possibly described by multiple words or word phrases, is called a "synonym set" or "synset". There are more than 100,000 synsets in WordNet, majority of them are nouns (80,000+). Image Net aims at proposing average 1000 images in each synset to be shown. Each image is controlled for each high quality concept and annotated by human. ImageNet proposes millions of categorized images for many concepts in WordNet hierarchy. LabelMe: It contains 50000 JPEG images (40000 are used for training and the other 10000 for testing). The size of each image is 256 256 pixels. The performed annotation is in two different file formats. One of the amount of tags is between [1.0,10]. 1.0 imp lies that the object in the image is similar to the extracted images. If no sample o f the object class can be found in the image o r d ifferent levels overlap each other, then the amount of the tag will be calculated as 1.0. Tiny Image: It contains images in size of 32 32 and they are created and saved as big binary files. 400Gb of free d isk space is needed for this data set. This data set enjoys two versions of function for reading image data including: (i) Load tiny images.m  p lain Mat lab function (no MEX), runs under 32/64 b its and loads images according to their numbers. Use this by default. (ii) readtinybigbinary.m  Matlab wrapper for 64b it M EX function. It is a bit faster and more flexib le than (i), but requires a 64bit machine
I.J. Intelligent Systems and Applications, 2016, 12, 2636
28
A Review of M ethods of Instancebased Automatic Image Annotation
T able 1. A number of general data image set used in image retrieving filed. Dataset
Some versions
Number of images
Categories
Other cases
ImageNet
ILSVRC
More than 14,200,000
>21800 Tags
Need to 400Gb disk space
10
50,000 training images & 10,000 testing images
CIFAR10 CIFAR
CIFAR100
60,000
100
Caltech101
More than 9,000
102
40 to 800 images per group
Caltech256
More than 30,600
257

YFCC
100,000,000
MIRFlickr
1,000,000
Oxford
More than 45,000
Concepts > 100
Tags are metadata and the text files are available
SIMPLIcity
1,000 10
Pictures & text description for the 10 categories butterfly
CALTECH
FlICKR
WANG WBIIS
10,000
LabelMe

More than 30,000
183
Size 256 × 256 pixels
Tiny Image

More than 79,000,000
> 75000Tags
Size 32 × 32 pixels
Scene397 SUN SUN2012
More than 131,000
More than 29,000
NORB
908 scenes Objects > 4400
6
T able 2. T he relationship between objects
Pictures toy s 6 general categories of animals, humans, ...
d b a
NUSWIDE
More than 269,600
81
One of the important criteria set annotations
SUNAttribute

More than 14,000
>700Tags

PascalVOC2007
VOC2005 – VOC2012
More than 9,900
20
To detect object class
c
a
(a = d < a = b < c , a = a < b = c < d)
(1)
C. The measures of similarity
B. Extracting the features of the images. In image retrieving systems the images are shown by the help of lo w level features since an image is a non structured array of p ixels. The first phase of semantic understanding is to extract applicable and effective v isual features of the pixels. To properly show the features creates a significant enhancement in semantic learn ing techniques. Both local and global showings are used in the techniques. The tendency is towards local features. In order to extract and calculate the local features the images need to be segmented, while the global features are extracted and calculated fro m the whole image. Image annotation mainly aims at finding the content of an image through extracted features. Some of the features are as follows:
397 Categories scenes
NUSWIDELITE NUSWIDEOBJECT NUSWIDESCENE
according to the pixels. The methods used for extracting the features of texture, t wo groups of space texture extract ing and the method of extracting the texture are spectral. The feature of shape: shape is considered to be the most impo rtant sign in determining and recognizing things for human beings in the real world. The ext racting methods of the features of the shape, two shape designbased extracting methods of the shape, and extracting the features of the shape are zonebased. In the shape designbased method the features of the shape are only calculated by edgs of the shape, while in the zonebased method the ext racting of the features is calculated according to the whole zone. The feature of Spatial relat ionship: This feature determines the location of the object in the image or its relat ionship with other objects. The relative locations such as left, right, down, up, and center are used in learning processes which are based on concepts. A two dimensional model is used in the relationship between objects as follows:
The feature of color: Co lor is one of the most important feature of an image which is defined as a special co lor space or a model. Color feature is extracted from an image or zones of an image. The feature of texture: One of the most important features of an image is its texture. As long as color is a feature of pixels the texture is calculated
Copyright © 2016 MECS
In order to retrieve images the queried image needs to be compared to the images of the database. The comparison is carried out between extracted features o f the queried image and the extracted features of the images of the dataset. To carry such comparison out a measure is needed which is known as the similarity measure. There is a group of similarity measure called distance measure. Generally, the construction of the vectors of feature determines the type of the d istance measure which is used in the co mparison process of the similarity. This calculation distance measure implies the similarity between the queried image and the images of the database. In order to reach the most precise and the best running, the annotation system needs to use the similarity measure which recognizes the similarities carefully. Some of these measure are as follows: ManhatanL1, Eucled ianL2, ChebyshevL∞, Hamming, Mahalanobis, Cosine, EMD, KL d ivergence, and J divergence. These measures different fo r their main features, limits, and range of applicability:
Minkowski distance: This is one the measure which is vastly used in retrieving systems. If n dimensional feature vectors of X and Y are (x1 , x2 , … , xn ) and (y 1 , y 2 , … , y n ), then Minkowski distance between X and Y will be defined as
I.J. Intelligent Systems and Applications, 2016, 12, 2636
A Review of M ethods of Instancebased Automatic Image Annotation
follows: n
1
d X, Y ( x i yi ) r r
(2)
i 1
Where r is considered to be the factor o f norm and it is always r 1 . If r 1 it is considered to be Manhattan measure, if r 2 it is Euclidean and if r it is Chebyshev. Mahalanobis distance: Consider the points A and B d istribution. Mahalanobis distance measure calculates the distance between A and B by calculating standard deviation of A fro m the average of B. if S is Covariance matrix and n dimensional feature vectors of X and Y are respectively (x1 , x2 , … , xn ) and (y 1 , y 2 , … , y n ), Mahalanobis distance between X and Y will be defined as follows: n
1 1 r
d X , Y ( xi yi S ) r
m
m
ij
n
min( wp , wq ) i 1
i
(6)
j
i 1
(3) m
n
m
n
EMD P, Q dij fij / fij
(4)
Hamming distance: Given a finite data space F with n elements, the Hamming distance d(x,y) between two vectors x, y F is the number of coefficients in which they differ, or can be interpreted as the minimal nu mber of edges in a path connecting two vertices of ndimensional space. In the CBIR system, the hamming distance used to compute the dissimilarity between the feature vectors that represent database images and query image. The fu zzy Hamming distance (D) is an extension of Hamming d istance for vectors with real values. Hamming d istance between X and Y is defined as follows: (n)
n
d X, Y  x i yi 
n
i 1 j 1
If r=2 and the result of Covariance matrix is the main matrix itself, it will be equivalent to Euclidean distance measure. But, if S is a diametric matrix, it will be equivalent to normalized Euclidean distance measure. Cosine distance: if n dimensional feature vectors of X and Y are respectively (x1 , x2 , … , xn ) and (y 1 , y 2 , … , y n ), the distance will be the Means angle between the vectors. Cosine distance between X and Y is defined as follows:
 X.Y  d X, Y 1 cosθ 1  X . Y 
Earth Mover distance: The EMD is based on the transportation problem fro m linear optimization which targets the min imal cost that can be paid to transform one distribution into the other. For image retrieval, th is idea is combined with are presentation scheme of d istributions which is based on vector quantization for measuring perceptual similarity. This can be fo rmalized in a linear programming problem as follows: P = {(p 1 , wp1 ),…,(p m , wp m)} is the first signature with m clusters, where p i is the cluster representative and wpi is the cluster weight; and Q = {(q 1 , wq1 ),…,(q n , wqn )} is the second signature with n clusters; and D = [d ij ] is the matrix of ground distance where dij is the ground distance between clusters pi and qj . To compute a flow F = [fij ], where fij is the flow between p i and q j , that minimizes the overall cost:
f
i 1
29
(5)
i 1
i 1 j 1
(7)
i 1 j 1
KullbackLeibler and Jeffrey divergence distance: Based on the informat ion theory, the KL divergence measures how inefficient on average it would be to code one histogram using the other one as codebook. Given two histograms H={h i } and K={ki }, where hi and ki are the histogram b ins, the KullbackLeibler (KL) divergence is defined as follows: m h d KL H , K hi log( i ) ki i 1
(8)
D. The K nearest neighbors After calculat ing similarity by similarity measure the K nearest visual neighbors of the queried image need to be obtained. In order to do so, the obtained amounts of each image, which is considered to be the amount of similarity, are arranged in ascending and K nu mber of them are selected as samples of visual neighbors of the queried image. These samples have the most similar with the queried image when compared with other images and they are suitable candidates for retrieving. E. Applying one of automatic annotation methods Automatic image annotation is carried out by the help of different algorithms. Neighbor voting, tag ranking, etc. are some examples that the most important of which will be analyzed through next parts.
Where if x i yi , then  x i yi  will be 0 an d if x i yi , then  x i yi  will be 1.
Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 2636
30
A Review of M ethods of Instancebased Automatic Image Annotation
T able 3. .Summarizes the types of distance measures and lists the main characteristics of each type. Measures
Main attributes
ManhattanL1
Less affected by outliers and therefore noise in high dimensional data.
EucledianL2 Allows
normalized and weighted features.
Chebyshev
Mahalanobis
Cosine
Limitations Yields many false negat ives because of ignoring the neighboring bins, and gives near and far distant components the same weighting.  Sensitive to the sample topology.  Does not compensate for correlated variables.
Equation
Usage/Domains
Equ. 2 (r=1)
 Computes the dissimilarity between color images.  e.g. fuzzy clustering
Equ. 2 (r =2)
T he most commonly used method, e.gkmeans clustering.
 Maximum value distance.  Induced by the supremum norm/uniform norms.
Does not consider the similarity between different but related histogram bins.
Equ. 2 (r =3)
 Quadratic metric.  Incorporates both variances andcovariances Efficient to evaluate as only the nonzero dimensions considered.
Computation cost grows quadratically with the number of features.
Equ. 3
Not invariant to shifts in input.
Equ. 4
Computes absolute differences between coordinates of a pair of objects, e.g. fuzzy cmeans clustering. Improves classification by exploiting the data structure in the space. Efficient for sparse vectors.
Hamming
Efficient in preserving the similarity structure of data.
Counts only exact matches.
Equ. 5
EMD
 Signaturebased metric.  T he ability to cluster pixels in the feature space.  Allow partial matching.
Not suitable for global histograms (few bins invalidate the ground distances, while many bins degrades the speed).
Equ. 6 , 7
KL divergence
 Asymmetric  Nonnegative
 Sensitive to histogram binning.
Equ. 8
III. A REVIEW OF MET HODS OF INST ANCE BASED AUT OMAT IC IMAGE ANNOT AT ION
Image retriev ing is carried out by two major methods including (1) text base image retrieving and (2) contentbase image retrieving. In order to retrieve an image based on text it needs to be annotated in a dataset. Image annotation process can be done both automatically and manually. In manual annotation process the images are annotated by experienced people. As the number of images in a web is fairly big and the data in browsers are massive this method is almost impossible to be carried out. Accordingly, automatic image retrieving methods are good alternatives. Annotation has got a significant potential influence on understanding and searching images. Huge data sets of images are the main problem of this method. Today, image annotation has become a vast research subject. Some automatic annotation methods in three tasks including Tag Assignment, Refinement and Retrieval will be analyzed in next sections. A. Tag Assignment
Znaidia et al [11]. presented method for tag suggestion using visually similar images is given in figure 2. It consists in two main steps: creating a list of “candidate tags” from the visual neighbors of the untagged image then using them as pieces of evidence to be combined to provide the final list of predicted tags. Given an untagged image I, we start by searching the k nearest neighbors using
Copyright © 2016 MECS
 Identifies the nearest neighbor relationships.  e.g Image compression, and vector quantization.  Useful metric between signatures in different spaces.  Robust against clutters and occlusions.  Efficient for clustering. Computes dissimilarity between distributions, e.g. texturebased classification
visual information (color,te xture). First, we compute a BOW signature for each neighbor based on local soft coding. Second, a sumpooling operation across the BOW of the k nearest neighbors is performed to obtain the list of “candidate tags” (the most frequent). Finally, basic belief masses are obtained for each nearest neighbour using the distances between this pattern and its neighbors. Their fusion leads to the list of final predicted tags. Verbeek et al [3]. proposed the weighted nearest neighbor for tag assignment as follows: yiwò1, 1 to denote whether concept w is relevant for image i or not. The probability that concept w is relevant for image i, i.e. p yiw 1 , is obtained by taking a weighted sum of the relevance values for w of neighboring training images j. Formally, is defined as follows:
p yiw 1 ij p yiw 1 j
(9)
j
(10) Where π ij stands for the weight of training image j when it is predicting the annotation process for I.J. Intelligent Systems and Applications, 2016, 12, 2636
A Review of M ethods of Instancebased Automatic Image Annotation
image i. to make sure that the distribution is carried out properly, require that ii 0 and j ii 1 . Each term P yiw 1  j is the prediction according to neighbor j in the weighted sum. Neighbors predict that image I has got the same relevance for concept w with probability 1Ɛ. The introduction of ǫ is a technique to avoid zero prediction probabilit ies when none of the neighbors j have the correct relevance value. The parameters of the model, wh ich they will be introduce and below, control the weights. maximizing the loglikelihood of predicting correct annotations for training images in a leaveoneout manner helps to estimate the parameters. Excluding each training image, i.e. by setting π ii =0, as a neighbor of itself must be taken into account. The aim is to maximize i ,w Ln P yiw . Vinlage status of Liberty
relationship of tag t in image X with the number of occurrence of t in the annotations of visual neighbors of X. This method introduces a unique user by limit ing the neighbors to create more votings. Each user has got mo re than one image in neighbors set. Additionally, counting process of the occurrence of tags needs to be carried out in advance. This method is as follows:
Tagvote x, t k
Alex compos for music status Liberty
status Liberty
Status of Liberty Paris Luxemburg Canon 450d
31
Status Liberty Ny c Strret Theater
t
k
nt s
(11)
Where n t is the number of tagged images with t tag in s set. K=1,000 Chen, Fan, and Lin et al [9,14], proposed TagFeature method [9] fo r tag assignment. Enriching the feature of the images by adding an additional tag feature to each image is the core idea of this method. A tag wo rd, which is composed of d', is the most number of frequent tags in s. Afterwards, a twoclass linear SVM classifier is trained by LIBLINEA R. The positive training set includes p tagged images in S, the same amount of minus samples of training are randomly selected fro m untagged images. The output of the classifier, wh ich is related to a special dimension in the tags of the image, is probably obtained by Platt scale. After adding tags and visual features, a feature is obtained by adding d+d' dimension. TagFeature x, t is obtained for t test tag by retraining an SVM classifier by the help of added features. Being linear, the classifier groups all support vectors into one vector then tries to classify a test image by the help of this vector. This process is as follows:
Tag Local Soft Coding
TagFeature x, t b x , x t
Candidate tags Status 5.7 Liberty 5.2 Monument 4.8 Paris 1 NewYork 1 Theater 1 ….
(12)
Where Xt is the total weight of all supporting vectors and b is the intercept. In order to create mean ingful classifiers tags having at least 100 positive samples are used. While d' is almost 400 [4,2] and p=500 and if the number of images for being tagged is more, a random vector samp le is carried out.
Belief Theory
B. Tag retrieving Predicted tags Status 0.99 Liberty 0.98 New York 0.97 Monument 0.95 Paris 0 Theater 0 …..
Fig.2. an example of tags assignment based on belief theory and local coding.
Li et al [10]. proposed TagVot algorith m for tag assignment. This method estimates the
Copyright © 2016 MECS
Liu et al [5], proposed twophase tag ranking algorith m for tag retrieving. Given an image x and its tags, the first step produces an initial tag relevance score for each of the tags, obtained by (Gaussian) kernel density estimation on a set of ñ=1,000 images labeled with each tag, separately. Secondly, a random walk is performed on a tag graph where the edges are weighted by a tagwise similarity. Then use the same similarity as in Semantic Field. Notice that when applied for tag retrieval, the algorith m uses the rank of t instead of
I.J. Intelligent Systems and Applications, 2016, 12, 2636
32
A Review of M ethods of Instancebased Automatic Image Annotation
its score, i.e.,
tag ranking algorithm. The term l / lx is a t iebreaker when two images have the same tag rank. Hence, for a g iven tag t, TagRan king cannot distinguish relevant images fro m irrelevant images if t is the sole tag assigned to them.
TagRanking x, t rank t l / lx where rank(t ) returns the rank of t produced by the
T able 4. A review of some sample based automatic image annotation methods in tag assignment.
Annotation method
local soft coding and belief theory
Providers
Znaidia et all.
Date & Location
April 16–20, 2013, Dallas, T exas, USA.
A Weighted Nearest Neighbour Model
Verbeek et all
March 29–31, 2010, Philadelphia, Pennsylvania, USA.
T agVote
Li et all
ACM XXX X, X, Article X (March 2015)
T agFeature
Chen , Fan , Lin et all
ACM XXX X, X, Article X (March 2015)
Gu illau min and Verbeek et al [2,3], proposed TagProp method. neighbor voting and distance parametric learn ing are used in this method. In this method a possible framework is proposed in which the probability of using neighboring images based on their rank or their weight according to their distance is defined. TagTop algorith m is as follows:
TagProp x, t .I ( x , t ) i
i
(13)
Where π j is a nonnegative weight indicat ing the importance of the jth neighbor xj ,and I(xj ,t) returns 1 if xj is labeled with t, and 0 otherwise K=1,000 and the rankbased weights, which showed similar performance to the distancebased weights Differ fro m Tag Vote that uses tag prior to penalize frequent tags. Tag Prop promotes rare tags and penalizes frequent ones by training a logistic model per tag upon TagProp x, t . The
use of the logistic model makes TagProp a modelbased method. Zhu et al [13], proposed graph voting. Graph voting is an oriented graph in which the nodes are annotated images by t tag in X. there e i, j òE ?
Copyright © 2016 MECS
T he aims of method 1. creating a list of “candidate tags” from the visual neighbors of the untagged image 2. using them as pieces of evidence to be combined to provide the final list of predicted tags Using positive and negative samples of training with assuming the most relevant test tag t. Given the weight for neighbors 1.estimates the relationship of tag t in image X with the number of occurrence of t in the annotations of visual neighbors of X 2. introduces a unique user by limiting the neighbors t o create more votings 1. enriching the features of the images by adding an additional tag feature to each image 2. In order to create meaningful classifiers tags having at least 100 positive samples are used
exists, if and only if image i is in Nk (i). X={x1 ,x2 ,…,xn ) is a set of feature vectors for all annotated images with t tag that xi ϵ Rd is the feature vector for ith image in X set and n is the number of annotated images by t tag. Nk (i) refers to the K nearest neighbors of i based on parameters like Euclidian distance or cosine. It is worth noting that for calculating Nk (i) not only annotated images by t are considered, but nonannotated image by t must be taken into account. The whole set of images is considered in order to find the K nearest neighbor set of Nk (i) for an image of i. Creating voting graph can be briefly described as follows: (1) For tag t, all annotated images having t tag are collected and used as the nodes of the graph. (2) the k nearest neighbors of Nk (i) are obtained for each j image in X set in the whole set. If each I image in X set appear in Nk (i), then there is an edge from vertex i to j. (3) the weight of W ij edge is set based on visual relevance between i and j. Visual relevance between two images is calculated by (Gaussian) kernel function with a parameter of ϭ diameter :
 xi x j 2 wij exp 2
(14)
I.J. Intelligent Systems and Applications, 2016, 12, 2636
A Review of M ethods of Instancebased Automatic Image Annotation
33
Makadia et al [1] used the K nearest neighbor for tag retrieving. This algorithm estimates the relationship of a tag by retrieving the first k nearest neighbor from S set based on visual distance d, then estimates the number of occurrence of the tag in the allocated neighbor tags. Knn is KNN x, t : kt . Kt is the number of
images with t tag in visual neighborhood of X.
Fig.3. A sample of tag retrieving by voting graph method. T able 5. A review of some sample based automatic image annotation methods in tag retrieving. Annotation method
Providers
Date & Location
T agRanking
Liu et all
ACM XXX X, X, Article X (March 2015)
T agProp
Guillaumin , Verbeek et all
ACM XXX X, X, Article X (March 2015)
Vote graph
Zhu et all
July 6–11, 2014, Gold Coast, Queensland, Australia.
k nearest neighbors
Makadia , Ballan et all
April 1–4, 2014, Glasgow, United Kingdom.
C. Tag refinement
Lee and Yong et al [17], proposed belief theory and neighbor voting for tag refinement in order to remove irrelevant tags among relevant ones. Let Ti is a set of allocated tags to i. Generally, Ti includes: (1) Relevant tags according to i content. (2) Irrelevant tags according to i content. During tag refinement if the relevance of a tag is less than a special threshold ξtag , T is irrelevant and Ti is removed:
Ti refined {t  t Ti r (t , i) ξtag}
(15)
Where Ti refined is a refined set of tags and ξtag determines if t is relevant or irrelevant according to i content. r(t,i)=rsimilar(t,i,k)rdissimilar(t,i,l) where rsimilar(t,i,k) denotes the relevance of t with respect to I when making use of k folksonomy images visually similar to i, and rdissimilar(t,i,l) denotes the Copyright © 2016 MECS
T he aims of method 1. produces an initial tag relevance score for each of the tags 2. Secondly, a random walk is performed on a tag graph where the edges are weighted by a tagwise similarity 1. uses neighbor voting and distance parametric learning 2. promotes rare tags and penalizes frequent ones by training a logistic model 1. Use annotate images and unannotated images 2. Creating voting graph can be briefly described estimates the relationship of a tag by retrieving the first k nearest neighbor from S set based on visual distance d, then estimates the number of occurrence of the tag in the allocated neighbor tags
relevance of t with respect to i when making use of l folksonomy images visually dissimilar to i. (1) Neighbor voting is used in order to estimate rsimilar(t,i,k). The relevance of t based on i content is estimated as the difference among "annotated images with t in a set of k retrieved neighboring images of i fro m ranked images by the help of visual similarity search" and "number of annotated images with t in a set of k retrieved neighboring images of i fro m ranked images by the help of" random sampling method. (2) Visual dissimilarity is used in order to estimate rdissimilar(t,i,l). The relationship of t according to i content as the difference between "annotated images with t in a set of l images which are dissimilar to I is estimated. L images of ranked ones are retrieved by the help of visual dissimilarity search" and "annotated images with t are estimated in a set of l neighboring images which are retrieved by the help of random sampling method". The lower I.J. Intelligent Systems and Applications, 2016, 12, 2636
34
A Review of M ethods of Instancebased Automatic Image Annotation
rdissimilar(t,i,l) is, the more t is related to i. i
Ti connecticut, clouds, coast,flower, grass, happy , leaves,mountain, pretty , rain, sad,sky , sun, trees, water, waterbury , 2009
Image folksonomy
Set of Visually similar images
Set of Visually disimilar images
Tag relevance estimation
Tag relevance estimation
Tag refinement
i flower, grass, leaves
Fig.4. A sample of tag refinement by belief and neighbor voting method
Van et al [15] proposed TagCooccur [7] method, which is based on tags, for tag retriev ing. This method used the test rank of the tag in the list of tags ranking. The list is created by ordering all tags when they occur the tag simu ltaneously. this method also account for the calculated stabletag count through its occurrence. Zhu et al [16] proposed RobustPCA method [6]. This method is based on analyzing the main powerful factors, D matrix (tag and image) is factorization by analysis of low Ran k decomposition with error scarcity and it is D=Ď+E in which Ď has a low rank constraint based on the nuclear norm, and E is an error mat rix with l1 
normsparsity constraint. Notice that this decomposition is not unique. The process of the image and tag nearness, as a solution, is carried out by adding two ext ra penalt ies with respect to a Laplacian matrix fro m the image affinity graph and another Laplacian matrix Lt fro m the tag affinity graph and it is relatively time consuming. Accordingly, two metaparameters ʎ 1 and ʎ2 are introduced in order to balance the error scarcity (two advantage of Laplacian). Two parameters are followed by a network search in the proposed area and a pretty stable algorith m is found. ʎ1 =20 and ʎ 2 =210 are empirically selected. As the users' tags are usually lost, the researchers has proposed preprocessing phase in which D is valued with weighing Knn propagation based on visual similarity. Truong et al [17], proposed “Tag InfluenceUnaware Neighbor Vot ing” method. In usual methods of voting all tags of the neighboring image of d' are supposed to have the same influence, according to the voting process, on describing visual contents of the image. For instance, relevance(tꞌ,dꞌ)=1, ⱯtꞌϵTdꞌ while the tags in a neighboring image tꞌϵTdꞌ have various applications in describing d'. However, it preferable to carry out the learning process for tag related to each d òD image. Afterwards, relevance(tꞌ,dꞌ)=1 (normalized[0,1]) is used for relearn ing related tag for a queried image. Notice that if the tag assignment compatib le and its relationship are considered together, the noise can be identified easily. For instance, a tag like tꞌϵTdꞌ may be a litt le related to d', but it is vastly related to t voting. Accordingly, this refinement process is considered as precise compatibility of the tag.
T able 6. A review of some sample based automatic image annotation in tag refinement. Annotation method
Providers
Date & Location
Belief T heory and Neighbor Voting
Lee , Yong et all
October 29, 2012, Nara, Japan
Van et all
ACM XXX X, X, Article X (March 2015)
Zhu et all
July 6–11, 2014, Gold Coast, Queensland, Australia.
analyzing the main powerful factors, D matrix (tag and image) is factorization by analysis of low Rankdecomposition with error scarcity
T roung et all
12, June 58, Hong Kong, China
T ag InfluenceUnaware Neighbor Voting method. In usual methods of voting all tags of the neighboring image of d' are supposed to have the same influence, according to the voting process, on describing visual contents of the image
T agCooccur
RobustPCA
T ag InfluenceUnaware Neighbor Voting
Copyright © 2016 MECS
T he aims of method belief theory and neighbor voting for tag refinement in order to remove irrelevant tags among relevant ones. the relevance of a tag is less than a special threshold test rank of the tag in the list of tags ranking. The list is created by ordering all tags when they occur the tag simultaneously. account for the calculated stable tag count through its occurrence.
I.J. Intelligent Systems and Applications, 2016, 12, 2636
A Review of M ethods of Instancebased Automatic Image Annotation
IV. CONCLUSION AND CHALLENGES In spite of previous works on instance based automatic image annotation, it is still considered to be a challenge in this field. In this paper, instance based automatic image annotation methods were reviewed. The main parts of automatic image annotation and various similarity measures were firefly discussed. Afterwards, instance based automatic image annotation methods were discussed in three fields including assignment, refinement and retrieving of tags. Being massive, the volu me of the images in the dataset made the annotation algorithm to be time consuming. Vo lu me and the number of created samples wh ich is followed by neighboring estimation is a big challenge. Each above mentioned methods has advantages and disadvantages and rely on some specific feature of the images and they are defined based on data center and specific application. Having the methods combined increases the efficiency since they present more informat ion about the image. Local features have a h igh differentiation power, but they are sensitive to noises and have less global differentiation power attributions and they are more stable than the noises. The most important challenges in instance based automatic image annotation are as follows:
The first challenge is to analyze the images with a high number of features. All features have limitat ions in interpreting the images and none of them can efficiently interpret the images of nature. Co mbin ing the features can be us eful, but to analyze them is very complicated. Accordingly, choosing an suitable number of features seems to be essential in image annotating. The second challenge is to create an efficient model of annotating. Most current models learn fro m low level features of the images, but the number of samp les for accurate train ing of a model is not big enough. Accordingly, texture informat ion and metadata need to be used in annotating. How to co mbine both low level v isual informat ion and high level texture informat ion together is a basic challenge. Today, annotation and online ranking are carried out simu ltaneously with several tags and they are not efficient enough in image retriev ing. The solution is to do annotation offline with monotag method then to rank the tag separately. In this method, first, the image is annotated then it is ranked offline. The fourth challenge is the lack of standard and classified words for annotation. Now optional words are used. Consequently, it is not still clear that how the image is grouped. A hierarch ical model of concepts is needed to accurately group the images.
Copyright © 2016 MECS
35
The next challenge is the weak tags of the images of the training set. Weak tag refers to tagged words and areas of the image that do not truly represent the content. For each image there are words that are tagged to the whole image and it is not clear which word refers to which area. REFERENCES
[1]
[2]
[3]
[4]
[5] [6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
A. M akadia, V. Pavlovic, and S. Kumar. 2010. Baselines for Image Annotation. International Journal of Computer Vision 90, 1 (2010), 88–105. M . Guillaumin, T. M ensink, J. Verbeek, and C. Schmid. 2009. TagProp: Discriminative M etric Learning in Nearest Neighbor M odels for Image AutoAnnotation. In Proc. of ICCV. J. Verbeek, M . Guillaumin, T. M ensink, and C. Schmid. 2010. Image annotation with TagProp on the M IRFLICKR set. In Proc. of ACM M IR. X. Li, C. Snoek, and M . Worring. 2009b. Learning Social Tag Relevance by Neighbor Voting. IEEE Transactions on M ultimedia 11, 7 (2009), 1310–1322. D. Liu, X.S. Hua, L. Yang, M . Wang, and H.J.Zhang. 2009. Tag Ranking. In Proc. of WWW. G. Zhu, S. Yan, and Y. M a. 2010. Image Tag Refinement Towards LowRank, ContentTag Prior and Error Sparsity. In Proc. of ACM M ultimedia. K. van de Sande, T. Gevers, and C. Snoek. 2010. Evaluating Color Descriptors for Object and Scene Recognition. IEEE Transactions on Pattern Analysis and M achine Intelligence 32, 9 (2010), 1582–1596. J. Sang, C. Xu, and J. Liu. 2012. UserAware Image Tag Refinement via Ternary Semantic Analysis. IEEE Transactions on M ultimedia 14, 3 (2012), 883–895. L. Chen, D. Xu, I. Tsang, and J. Luo. 2012. TagBased Image Retrieval Improved by Augmented Features and GroupBased Refinement. IEEE Transactions on M ultimedia 14, 4 (2012), 1057–1067. X. Li and C. Snoek. 2013. Classify ing tag relevance with relevant positive and negative examples. In Proc. of ACM M ultimedia. A. Znaidia, H. Le Borgne, and C. Hudelot. 2013. Tag Completion Based on Belief Theory and Neighbor Voting. In Proc. of ACM ICM R. Z. Lin, G. Ding, M . Hu, J. Wang, and X. Ye. 2013. Image Tag Completion via ImageSpecific and TagSpecific Linear Sparse Reconstructions. In Proc. of CVPR. X. Zhu, W. Nejdl, and M . Georgescu. 2014. An Adaptive Teleportation Random Walk M odel for Learning Social Tag Relevance. In Proc. of SIGIR. Y. Yang, Y. Gao, H. Zhang, J. Shao, and T.S. Chua. 2014. Image Tagging with Social Assistance. In Proc. Of ACM ICM R.. K. van de Sande, T. Gevers, and C. Snoek. 2010. Evaluating Color Descriptors for Object and Scene Recognition. IEEE Transactions on Pattern Analysis and M achine Intelligence 32, 9 (2010), 1582–1596. G. Zhu, S. Yan, and Y. M a. 2010. Image Tag Refinement Towards LowRank, ContentTag Prior and Error Sparsity. In Proc. of ACM M ultimedia. B. Truong, A. Sun, and S. Bhowmick. 2012. Content is still king: the effect of neighbor voting schemes on tag relevance for social image retrieval. In Proc. of ACM ICM R.
I.J. Intelligent Systems and Applications, 2016, 12, 2636
36
A Review of M ethods of Instancebased Automatic Image Annotation
Authorsâ€™ Profiles Morad Derakhshan born in 1972 in Kamyaran. Now live in Sanandaj, Iran. His job Employee Education Kurdistan and Graduate student of Software, Dep artment of computer Engineering, Sanandaj Branch, Islamic Azad University, Sanandaj , Iran. His interest field of research, Image Processing, Evolutionary Algorithms and Data M ining
Vafa Maihami born in 1987 and received the M .S. degree in computer engineering from the Kurdistan University, Sanandaj, Iran, in September 2012. Since 2010 he is lecturer in Sanandaj Branch of Islamic Azad University and other institutes in Sanandaj. Currently he is facility member at Sanandaj Branch of Islamic Azad University. His interest field of research Image Processing , Computer Vision , Information Retrieval , M achine Learning , Wireless Sensor Network.
How to cite this paper: M orad Derakhshan, Vafa M aihami, "A Review of M ethods of Instancebased Automatic Image Annotation", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.2636, 2016. DOI: 10.5815/ijisa.2016.12.04
Copyright ÂŠ 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 2636
I.J. Intelligent Systems and Applications, 2016, 12, 3745 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.05
A Survey on Speech Enhancement Methodologies Ravi Kumar. K Lakireddy Balireddy Engineering College, Mylavaram, 521230, India Email: 2k6ravi@gmail.co m
P.V. Subbaiah V. R. Siddhartha Engineering College, Vijayawada, 520001, India Email: pvsubbaiah@vrsiddhartha.ac.in
Abstractâ€”Speech enhancement is a technique wh ich processes the noisy speech signal. The aim of speech enhancement is to improve the perceived quality of speech and/or to improve its intelligib ility. Due to its vast applications in mobile telephony, VOIP, hearing aids, Skype and speaker recognition, the challenges in speech enhancement have grown over the years. It is more challenging to suppress back ground noise that effects human co mmunicat ion in noisy environ ments like airports, road works, traffic, and cars. The object ive of this survey paper is to outline the single channel speech enhancement methodologies used for enhancing the speech signal wh ich is corrupted with addit ive background noise and also discuss the challenges and opportunities of single channel speech enhancement. This paper main ly focuses on transform do main techniques and supervised (NMF, HMM) speech enhancement techniques. This paper gives frame work for developments in speech enhancement methodologies. Index Termsâ€”W iener filter, Bayesian Estimators, Super Gaussian priors, Nonnegative Matrix Factorizat ion (NMF), Hidden Markov Model (HMM), Phase Processing.
I. INT RODUCT ION In human co mmun ication speech signal adversely affect by additive background noise and the speech may be degraded. The degraded speech is uncomfortable for human listening and hence the degraded speech must be processed. In speech communication system generally noise is reduced at farend to improve the quality and at nearend speech modification is done to improve th e Intelligib ility. Thus speech enhancing/processing aims to improve quality of the speech and/or intelligib ility of the speech. To simplify speech enhancement problem necessary assumptions about the nature of noise signal must be considered. Generally noise can be assumed as additive, stationary noise. Over the years researchers developed significant approaches to enhance the corrupted speech and obtained satisfactory improvements under high SNR conditions. Conventional algorithms [13] like spectral subtraction, wiener filtering, and subspace Copyright ÂŠ 2016 MECS
approach enhances the corrupted speech and poses the limitat ions like musical noise. In the process of enhancing, the enhancers attenuate some co mponents of speech and results in intelligibility reduction [3], i.e., if quality are achieved using some methodology there is effect on intelligib ility due to processing. Similarly the improvement in intellig ibility poses reduction in quality of speech. Hence techniques like processing in modulation do main [4] came in to p icture for t radeoff between intelligibility and quality [3]. The foremost thing that has to know is what causes the speech to degrade. The degradation may cause due to unnoticeable background noise, degradation results fro m mu ltip le reflect ions, [3] and using inappropriate gain. The papers [15] provides methodologies on how to enhance the speech when speech is degraded by additive back ground noise, as it has many useful applicat ions in daily life like using mobile in noise environ ments like offices, cafeteria, busy streets and web applications like Skype, Gtalk and sending commands fro m cockpit of aero p lane to ground. To overcome this, many speech enhancement techniques are using noise estimation as its first step [12]. Noise estimation can be done mostly in spectral do main like spectral magn itudes, spectral powers. Another approach is using voice activity detector. Vo ice Activ ity Detector (VA D) estimates the noise during speech pauses and averages the signal power during all these intervals. Also one important thing while processing the speech in frequency domain is that the processing is done by dividing the speech in to overlapping frames using Hanning / Hamming window and then ShortTime Fourier Transform (STFT) is applied. This step is necessary as speech itself is nonstationary and the transform techniques work only for stationary signals. To apply signal processing techniques the speech is considered to be short time stationary [6] and hence framing must be done. But, some speech enhancement techniques like [3] Adaptive filters, co mb filters, kalman filters are processed in time do main. In such case methods applied direct ly on speech signal itself. The main goal of speech enhancement is either to improve quality or intelligibility or both, that depends on the type of applicat ion. Fo r hearing impaired listeners , the main criterion is intelligib ility improvement. Th is can be done I.J. Intelligent Systems and Applications, 2016, 12, 3745
38
A Survey on Speech Enhancement M ethodologies
by frequency compression and bandwidth expansion. That is why noise reduction can be seen as speech restoration and speech enhancement. In speech restoration, the degraded speech can be restored as original speech where as in speech enhancement it tries the processed signal to be better than unprocessed signal. Thus both terms can be used interchangeably. Estimators using Gaussian and SuperGaussian prior are developed for better noise reduction [714]. As the research is going on, researchers used the perceptual [3, 6] properties of human ear like masking of inaudible co mponents and are successful in obtaining improved results. Supervised methods [1519] uses training of noise and speech samples and hence no further requirement of VAD calculation and hence obtain improved results. Up to some decades, researchers process the amplitudes of noisy speech where as noisy phase is being unprocessed, as human air is insensitive to phase information. Later on researchers find that phase informat ion is useful [1923] under low SNR cases. Now res earchers shown that, the performance of the speech enhancement will improve by processing the noisy signal phase along with the noisy Amplitudes. The paper is organized as follows: Section II gives single channel speech enhancement methodologies; Section III d iscusses the transform do main approaches, Section IV p rovides decomposition techniques like NMF, Modeling methods using HMM and also provides the significance of phase processing. Section V prov ides the
challenges and opportunities in single channel speech enhancement methodologies. Section VI gives conclusion.
II. SINGLE CHANNEL SPEECH ENHANCEMENT M ET HODOLOGIES Single channel speech enhancement problem mainly deals with the applicat ions where a single microphone is used for recording purpose such as mobile telephony. These techniques provide improvement in the quality o f degraded speech. Basically all the methods classify into two categories, one is supervised methods and the others are unsupervised. In supervised methods like NMF, HMM, noise and speech are modeled and parameters are obtained using training samples [16]. Whereas in unsupervised methods like transform domain approaches given in Fig.1, Wiener filter, Kalman filter and estimators using SuperGaussian without knowing prior informat ion about speaker identity and noise, processing is done [13]. Supervised methods do not require the calcu lation of Noise Power Spectral Density (PSD), which is one of the difficult tasks in speech enhancement. For better understanding of developments in speech enhancement methods, they are classified as shown in Fig.1. Generally frequency domain processing [3] is easy and more understandable and hence transform do main methodologies play predominant role. Classification of speech enhancement methodologies are given in Fig.1
Single channel Speech Enhancement
Time Domain Processing 1) 2) 3)
Phase + Amplitude Processing
Kalman filter Adaptive filters Intelligent filters (Using optimization)
Supervised Techniques
Transform Domain 1) 2) 3)
4) 5)
Decomposition Based
M odelling of Speech/ Noise
Spectral Subtraction 1) NM F Wavelet Transform Stochastic Estimation 3.1) Wiener Filter (linear Estimator) 3.2) Non linear Estimators (For Super Gaussian priors) a) M M SE b) M M SELSA c) M AP d) M L Super Gaussian Estimates ( + STFT phase) M odulation Domain
1) HM M 2) GM M
Fig.1. Speech Enhancement Methodologies
Copyright ÂŠ 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 3745
A Survey on Speech Enhancement M ethodologies
T able 1. Spectral Subtraction Methodologies
III. TRANSFORM DOMAIN A PPROACHES In transform do main approaches, signal is converted into frequency domain and processing is done on DFT coefficients/Wavelet coefficients. The advantage is that it is easy to identify noise and speech components and hence noisy components can discard. Different estimators are developed (uses transform do main processing) for speech enhancement using Gaussian and SuperGaussian Speech priors. Speech Presence Uncertainty is taken in to consideration for better performance in removing residual noise. Let y(n) be the input noisy signal and b(n) be the noise and x(n) be clean speech signal. Assume noise is addit ive then we write y(n) x(n) b(n)
a) Basic Spectral Subtraction
Sˆ()
2
Y ()
2
Bˆ ()
2
b) Spectral Oversubtraction
Sˆ ()
2
2 Y ( ) 2 ˆ B ( )
Bˆ ()
2
if
Y ()
else
2
2 ˆ () ( ) B
c) Multiband Spectral Subtraction 2 2 2 ˆ i() if Sˆ i() Y i() 2 2 Y i( ) i i B Sˆ i() 2 else Y i()
; k k i
i 1
(1)
In frequency domain p rocessing (STFT) done on frames, due to discontinuities at frame boundaries there introduces some distortion [3]. This can be controlled by the choice of windowing function and ratio of frame increment to window length. Best choice is using Hanning window taking frame increment 2 and window length 2 for perfect reconstruction [13] and to attenuate discontinuities at frame boundaries. For the case of Hamming window, the ratio 4 is used. Analysis of window shift, based on pitch period works better for frequency domain processing. A. Spectral Subtraction Spectral Subtraction is one of the first and significant approaches for single channel speech enhancement [14]. In spectral subtraction, estimated clean speech is obtained by subtracting the estimated noise spectrum fro m the noisy signal (clean speech + noise) and results in estimated clean speech. Spectral subtraction suffers fro m remnant noise, musical noise and speech distortion [12]. To address this problem, several variations of the basic method are proposed. One variat ion is taking the control over amount of noise power subtracted from noisy power spectrum (spectral over subtraction) [3]. Here a constant subtraction factor is used for the total length of the spectrum and results in significant amount of reduction in remnant noise. For further reduction of remnant noise, another approach is that, firstly use basic spectral subtraction and obtains the enhanced speech. After that, again give the enhanced output as input and iteratively repeat the process for number of t imes gives better reduction in remnant noise (Iterative spectral subtraction). In real world, noise affects the speech differently for different frequencies. To deal the real world noise, rather than using constant subtraction factor for whole spectrum, the subtraction factor is set individually for each frequency band (Multiband spectral subtraction). At low SNRs the subtraction factor is unable to adapt with variable noise characteristics [13]. By using masking properties of human auditory system, attenuate the noise components that are not audible due to masking (spectral subtraction using human auditory system. Copyright © 2016 MECS
39
Afterwards spectral subtraction was implemented in modulation domain. Here the subtraction is performed on the real and imag inary spectra separately, in modulation frequency domain (so it enhances the magnitude and phase). B. Wavelet transform General DFT approach is not able to localize time and frequency, i.e., it is not able to provide the exact frequency at exact time. Whereas the wavelet transform provides good time frequency analysis [10]. Hence using wavelet transform, it is able to know the frequency information at a particular time. Wavelet uses variable time windows for different frequency bands and hence better resolution is achieved at low frequency bands as well as high frequency bands. Wavelet transform is a powerful tool to deal speech signals which are normally nonstationary and hence used for noise reduction in single channel speech enhancement. After applying the wavelet transform, the coefficients can be modified by putting some threshold value such that the noise coefficients can be neglected and hence noise reduction is possible. To achieve better performance in single channel speech enhancement, where only one microphone is used, the subband processing approach is worthwhile. Human auditory system is divided in to 18 critical bands [10] (frequency bands) and process the signal according to these critical bands yield better results. C. Stochastic Estimation 1) Wiener Filter Wiener filter suppress the noise by min imizing the Mean Square Error (MSE) between the original signal magnitude and the processed signal magnitude. H (k )
Pxx ( ) Pxx ( ) Pbb ( )
(2)
Where Pxx ( ) , Pbb ( ) are the clean signal power spectra and noise power spectrum respectively [3, 10]. It is observed that under low SNR conditions the ratio becomes very small and approaches to zero, i.e., I.J. Intelligent Systems and Applications, 2016, 12, 3745
40
A Survey on Speech Enhancement M ethodologies
H (k ) 0 and at extremely h igh SNR regions the ratio
approaches to unity, i.e., H (k ) 1. Hence Wiener filter attenuates under low SNR and emphasizes under high SNR. Later so me parameters are added to the Wiener filter to achieve d ifferent characteristics at different SNRs as Pxx ( ) H (k ) P ( ) P ( ) bb xx
(3)
Where , are the parameters used to obtain different attenuation characteristics for different values. It is noted that the above Wiener filter is noncausal as it requires knowledge of clean signal. Iterative W iener filter is used for estimat ion in iterat ive fashion. Afterwards , some spectral constraints are imposed within frame and are considered for processing and constrained Wiener filtering algorith m is proposed. In subband Wiener filter, the signal is divided according to human auditory critical bands and Wiener estimation is applied on each band. Also while processing it is useful to vary the window size [10] according to pitch. Later perceptual constraints are introduced in Wiener filter and masking of inaudible sources is incorporated with Wiener filters. 2) Non linear Estimators Wiener filter assumes the DFT coefficients of both speech and noise as Gaussian random variables (STFT). Wiener filter is linear estimator as it estimates comp lex spectrum with MMSE. But in M L and MMSE estimators, estimation of modulus of DFT coefficients is done which is a non linear estimation process. It is noted that Wiener estimator is optimal as co mplex magnitude spectral estimator [3, 7, 8]. In Wiener filter, mean is calculated with X (k ) rather than X k . Bayesian estimators are proposed with ShortTime Fourier Transform (ST FT) coefficients as well as with ShortTime Spectral Amplitude (STSA). In STSA, estimat ion of spectral amp litudes is done where as in STFT; estimation of complex spectrum is done. i. Minimum Mean Square Error (MMSE) Estimator Quality and intellig ibility can improve if the estimator is optimu m in spectral amplitude sense, i.e., minimizing the mean square error of short time spectral amplitude of processed and true magnitudes [78].
e E Xˆ k X k
2
(4)
Where Xˆ k and X k are the magnitudes of processed and clean speech respectively. Calcu lation of MMSE estimator gain [11], requires the knowledge of Bessel functions. It also requires lookup table.
vk
k
(1.5) M (0.5,1; vk )
Copyright © 2016 MECS
Where vk
k k , function of priori SNR and 1 k
posteriori SNR and M (0.5,1; vk ) is hypo geometric function. However, co mputationally efficient techniques without usage of Bessel functions are proposed. This can be achieved by either Maximu m A Posteriori or MMSE estimation of spectral power. ii. MMSE LogSpectral Amplitude Estimator (MMSELSA) Mean square error or cost function is included with logarithm function as
e E log X k log Xˆ
2
(6)
The Gain for MMSELSA Estimator is given as [3, 11] t 1 e exp dt k 2 t vk
vk
(7)
This estimator results in a slightly higher speech distortion but lower residual noise than MMSE STSA (due to higher suppression i.e. smaller gain). It suffers fro m less residual noise than MMSE Estimator [3, 7, 11] and also maintains the quality of enhanced speech same as obtained with MMSE (even by taking Speech Presence Uncertainty (SPU) in to account). Note that SPU in log STSA is unworthy. All the estimators are generalized using a cost function with different values of α, β, as [11]
C ( K ˆ K )2 ( K2 K2 )
(8)
Different values of α, β, results in different estimators. Gains of each estimator is used for obtaining enhanced speech using (Gain multiplied with noisy speech)
Xˆ Gain.Noisy
(9)
One of the estimators is βorder MMSE estimator with gain given as
vk 1 M ;1; vk k 2 2
(10)
This estimator provides better tradeoff between residual noise and speech distortion. Better results can achieve if appropriate stochastic parameters are used for β value adaptation like masking threshold. Decreasing β below 0 causes an increase in the noise reduction and speech distortion. β estimator with β = −1 slightly shows better performance than MMSE STSA and LSA estimators in terms of PESQ [7,11,13]. Weighted Euclidean (WE) Estimator gain can be calculated as
(5)
I.J. Intelligent Systems and Applications, 2016, 12, 3745
A Survey on Speech Enhancement M ethodologies
vk
k
p 1 p 1 1 M ,1; vk 2 2 p p 1 M ,1; vk 2 2
(11)
Under high SNR conditions all the above mentioned estimators approaches to wiener estimator [11]. iii. Maximum A Posteriori (MAP) Estimator In MAP approach choose
ˆ arg max P( / x)
(12)
And
p( / x)
p( x / ) p( ) p ( x)
(13)
Here MAP Estimator does not depend on p(x) and the MAP estimator is given as [3, 14]
ˆ arg max ln( p( / x)) ln( p( )
MAP estimators of the magnitudesquared spectrum is obtained as
P(YK H1k ) P( H1K ) P(YK H ) P( H 0K ) P(YK H1k ) P( H1K ) k 0
(16)
Where H1 is for speech presence hypothesis and H0 is for speech absence hypothesis. Multiply the estimator gain with SPU probability to obtain the enhanced speech. D. Modulation Domain
2 (1 ) 2(1 )
length durations (<25ms) rather than using Gaussian assumption for speech DFT coefficients Laplacian and Gamma (SuperGaussian) assumption yields better results as the variance is less for superGaussian (Laplacian, Gamma) than Gaussian . Also using superGaussian ones can obtain distribution with sharp peaks and less tails than Gaussian as shown in Fig.2. Hence researchers proposed estimators using superGaussian speech priors. Use of super Gaussian speech priori g ives better noise reduction but with poorer nois e equality, i. e, increased noise reduction and significant rise in distortion. SuperGaussian priors do not exactly match with the distributions measured and hence further improvements in noise reduction is still possible by considering better spectral variance estimators and alternate PDF models. Some researchers have shown that using Speech Presence Uncertainty along with estimator g ives better performance (to deal residual noise). The Speech Uncertainty probability is given by P( H1k Y (k ))
(14)
41
(15)
The above mentioned MAP estimator is a powerfu l tool to improve speech intellig ibility, at ext remely low SNR level [14]. In some estimators, the DFT coefficients of speech and noise are assumed as Gaussian probability density function (pdf) and also several works used nonGaussian speech priors and better results are obtained [12,14].
Intelligib ility is a key factor for understanding good percentage of words even under noisy conditions (reduces listening effort). At low SNRs normal people can understand more words than patients with hearing impaired [4]. Intelligib ility can be possible if there is significant reproduction of modulation of the spectral amp litudes. The processing of the signal in modulation domain is given in Fig. 3 Noisy Speech
iv. Estimators Using SuperGaussian Speech Priors Obtain Spectrogram
S (t , )
Obtain Time Trajectory and Remove Mean
S (t , k )
Spectral Analysis
Obtain Fig.2. Super Gaussian Distribution  Lesser Variance and High Spectral Peaks
Under Gaussian speech and noise priors, W iener estimator is regarded as optimal. But for lesser frame Copyright © 2016 MECS
S ( , k ) and S ( , k ) Modulation Spectrum Fig.3. Processing of Noisy Speech in Modulation Domain
I.J. Intelligent Systems and Applications, 2016, 12, 3745
42
A Survey on Speech Enhancement M ethodologies
IV. NMF AND HMM A PPROACHES Speech enhancement methods like, Wiener filter, Spectral Subtractions are not able to give satisfactory results for nonstationary signals. To achieve significant improvement in processed speech (quality) noise typ e must be known in advance. To deal this, enhancement methods based on HMM and NMF were used. In these methods noise and speech samples have to be trained. A. Speech Enhancement Using NMF NMF based speech enhancement is a decomposition technique useful for denoising and especially for denoising nonstationary signals like speech. In this technique, decomposition of signal is done as a combination of nonnegative building blocks. Optimal choice of W and H are obtained by solving V WH [15, 17]. In speech enhancement problem, it is considered that, the signal V as spectrogram and the building blocks, W as set of specific spectral shapes and H as activation levels respectively. Generally researchers us e Kullback Leibler (KL) d ivergence as one of the main object ive function. Variants of NMF can be obtained by choosing different objective functions. Three impo rtant phases of processing the signal are 1) Training phase 2) Denoising phase 3) Reconstruction phase. In training phase, Nonnegative matrix factorizat ion is performed on the clean speech and noise (Assume availability o f spectrograms of clean speech and noise), minimizing KL divergence between Vspeech and Wspeech H speech and also between Vnoise and Wnoise H noise . And the mean and variance for noise and speech H blocks are computed. In denoising, fix W blocks and find H, so that it minimizes the KL divergence [15]. Later, update blocks using any class of NMF algorith ms and finally reconstruct the enhanced
spectrogram. NM F can be implemented as supervised and unsupervised. B. Supervised Speech Enhancement using NMF Let Y, S, B be matrices of complex DFT coefficients of noisy, clean speech, noise. The Nonnegative transformation for Y, S, B is obtained and those are given as V, X, U such that
vkt ykt , xkt skt , ukt nkt , p
p
p
where p=1 and 2 for magn itude spectrogram and power (Magnitude square) spectrogram. In supervised approach, prior to enhancement basis matrix for speech, noise has to be learnt. Let it be Wspeech and Wnoise and obtain combined basis vector as
W WspeechWnoise
(17)
And noisy matrix obtained as (W is fixed)
vt W ht WspeechWnoise ht( s ) ht( n ) T
T
T
(18)
Finally enhanced speech is obtained using
Xˆ t
W ( s)ht( S ) .vt W ( s)ht( S ) W (n)ht( n )
(19)
as Wiener gain. The advantage with supervised approaches is that, no need of finding noise power spectral density and hence these approaches gives better results in enhancement process even for nonstationary noise [17].
Fig.4. Block Diagram for Supervised Speech Enhancement Methods (NMF/HMM)
C. Unsupervised Speech Enhancement using NMF Here the noise basis vectors are learned during intervals of speech pauses. By keep ing noise basis constant during speech activity period and parameters like speech basis, NMF coefficients of speech and noise are learned [17]. This learning is done by minimizing Euclidean distance and enhanced speech is obtained using Copyright © 2016 MECS
the above Wiener relation. D. Different Classes of NMF Classes of NMF algorith ms are Mult iplicative Update Algorithm, Grad ient Descent Algorithm, and Alternating Least Squares Algorithm 1) Multiplicative Update Algorithm
I.J. Intelligent Systems and Applications, 2016, 12, 3745
A Survey on Speech Enhancement M ethodologies
1) 2) 3)
Initialize W as random dense matrix Initialize H as random dense matrix Update W, H up to several iterations
Where 4) 5)
H H .*(W T A). / (W TWH )
(20)
W W .*( AH T ). / (WH H T )
(21)
is used to avoid zero, =10 9
Repeat step 3 for several iterations End
43
the sample vectors for noisy, clean speech and noise signals. This can be done by framing each of the signals and for each frame one sample vector is obtained. Let the noisy signal be yk sk bk . No w noisy signal is div ided into frames and are stored in vector (fo r each frame) as yT yk , yk 1 ,.... yk L 1 , where T denoting the frame index. In similar manner, obtain the vectors sT and bT . The speech signal and noise signal are first modeled using HMM and each of its output density as Gaussian Mixture Model (GMM ) [18]. Here speech and noise process is Auto Regressive (AR) and the hidden state probability for speech is given as s1K s1 ,....sK
2) Gradient Descent Algorithm K
1) 2) 3)
f ( s1K ; ) ... azT 1 zT f ( sT zT ; )
Initialize W as random dense matrix Initialize H as random dense matrix Update W,H up to several iterations
f H H H H
Q1
(27)
QK T 1
Where az0 z1 P( z1 ) is init ial state probabilities, here (22)
Z indicates states and the equation I
W W W 4) 5)
f W
(23)
matrix o f state z. And the co mbined HMM is obtained using Noise HMM and Speech HMM [19]. For finding the parameters, Expectation Maximizat ion (EM) algorith m is used. Nonnegative HMM is developed and implemented with MMSE estimator. Later, it was shown that better performance can ach ieve if HMM is co mbined with super Gaussian priors.
Initialize W as random Matrix Set all negative elements in H to zero. Solve for W using
HH T W T HAT 4) 5) 6)
(24)
Set all negative elements in W to zero Repeat step 2 to step 5 for several iterations End
4) Constrained NMF 1) 2) 3)
4) 5)
Initialize W as random dense matrix Initialize H as random dense matrix Update W, H up to several iterations
H H .*(W T A). / (W T WH H )
(25)
W W .*( AH T ). / (WHH T W )
(26)
Repeat step 3 to for several iterations end
E. Hidden Markov Model (HMM) Based Method Markov process is a stochastic model used to model a random system that changes states according to some transition rule wh ich depends on current state (only). HMM is a model which relates hidden variables and mixtu re weights through Markov process [19]. Let’s see how HMM is applied to speech enhancement. First obtain Copyright © 2016 MECS
(28)
i 1
is GMM which depends on state. Ci , zT is the covariance
Repeat step 3 to for several iterations End
3) Alternating Least Squares Algorithm 1) 2) 3)
f ( sT zT ; ) wi , zT g ( sT ;0, Ci , zT )
F. Phase Processing + Amplitude Processing Hu man ears are insensitive to phase informat ion. But later researchers came to know that phase is important factor for Intelligib ility [3]. In [21], it is showed that signal reconstruction is possible using phase only reconstruction. By co mbining processed phase with amp litude estimators, better enhanced speech may be obtained. From perceptual point of view, at high SNR noisy speech phase is close to clean speech phase and hence the noisy phase is used to replace clean phase. However, when SNR d rops low, noisy phase shows a negative effect and it might be perceived as “roughness” in speech quality [3, 21], i.e., even for clean magnitude spectrum at low SNR there is inability to recover clean speech with unperceivable distortion. At the early research time, researchers observed the magnitude spectrogram and phase spectrogram and came to a conclusion that spectral and temporal informat ion obtained by phase spectrogram is insignificant (due to phase wrapping) when compared to the informat ion obtained by magnitude spectrogram. Later researchers showed that by using group delay plot (derivative of phase with frequency) and instantaneous frequency plot, enough information about the speech signal can be obtained. Interestingly same information obtained by magnitude spectrogram can be obtained using derivatives I.J. Intelligent Systems and Applications, 2016, 12, 3745
44
A Survey on Speech Enhancement M ethodologies
of spectral and temporal phases. If phase systems are minimu m or maximu m, Hilbert transform is used to relate logmagnitude and phase which means either only the spectral phase or the spectral amplitude is required for signal reconstruction. But for the signals like speech , maximu m/ minimu m phase is restricted. It is noted that the SNR obtained when noisy magnitudes mixed with phase which is less distorted results in SNR improvements up to 1 dB. The STFT magnitude spectrum is important than phase spectrum, for segment length between 5 ms to 60 ms, and for segments which are shorter than 2 ms and longer than 120 ms, the phase spectrum plays crucial role. In contrast to this, signal segments of 32 ms length [21], overlap of 7/8th (Rather than 50%) during the STFT analysis , along with zero padding, the performance of magnitudebased speech enhancement can be significantly improved if processed phase is taken into account. The first and foremost approach is GL iterative approach (Griffin and Lim) [21]. In GL approach, updated phase information is retained where as the updated magnitudes are replaced. Later (Real Time Iterative Spectrogram Inversion) RTISILA is developed in which phase is updated in multiple frames. In sinusoidal modelbased phase estimat ion, fundamental frequency is used for estimating the clean spectral phase which is taken fro m the degraded signal. Each of these techniques has different difficult ies. Hence enhanced spectral magnitudes combine with processed phases can overcome these limitations. In [22], authors derived phase aware magnitude estimator based on MMSE estimator. One phaseaware co mplex estimator is the Co mp lex estimator with Uncertain Phase (CUP) [23]. The init ial phase estimation can be done using signal characteristics. The open issue is phase estimation is difficult at very low SNRs. This may overcome by jo ining the different phase processing approaches into iterative phase estimat ion approaches. In addition, better performance y ields by considering speech spectral coefficients as Gamma distributed and noise spectral coefficients as Gaussian [23]. Clean speech phase estimation is an interesting field of research in area of speech enhancement.
V. CHALLENGES AND OPPORT UNIT IES Speech enhancement objective is to improve quality and intellig ibility. Existing methods are not able to improve both quality and intelligib ility and tradeoff between the quality and intelligib ility is always needed. Develop ment of methods which provides less distortion while processing the speech is needed. Assuming speech priors as SuperGaussian in d ifferent estimators improved the performance of estimators but still these distributions not exactly match with speech DFT coefficients. There is a need for sophisticated speech priori assumptions. Under High SNR condit ions available speech enhancement methods are providing better results, but at low SNR conditions there is necessity to develop improved techniques. To deal with nonstationary signals like speech, there is need to develop supervised methods Copyright ÂŠ 2016 MECS
using NMF and HMM. Better results are obtained if statistical estimators are used along with NMF and HMM. Use of SuperGaussian in NMF and HMM may also lead to new speech enhancement methodologies. Unsupervised and Supervised methods ignored phase informat ion or phase processing due to its complexity. Joint A mplitude and Phase Estimation methods will place significant position in speech enhancement field. Amplitude estimators combine with processed phase informat ion will open new techniques in the field of speech enhancement.
VI. CONCLUSION In this paper, different speech enhancement methodologies and its developments are discussed. Bayesian Estimators and Frequency domain approaches plays significant role in noise reduction. Using speech presence uncertainty along with estimators can imp rove the performance of Estimators. Supervised methods like NMF and HMM are helpful fo r dealing Nonstationary signals. SuperGaussian Estimators included in NMF gives better noise reduction. There is need in considering processed phase informat ion along with amplitude information. REFERENCES [1]
Berouti, M . Schwartz, R. M akhoul, Enhancement of Noisy Speech Corrupted by Acoustic Noise, Proc.. of ICASSP 1979, pp.208211. [2] Boll,S.F. Supression of Acoustic Noise in Speech Using Spectral Subtraction, Proc.. of IEEE Trans AASP, Vol 27, No.2, 1979, pp. 113120 [3] P.C. Loizou, speech enhancement: Theory and practice, CRC press, 2007. [4] Kuldip Paliwal, Kamil Wojcicki, Single Channel Speech Enhancement Using Spectral Subtraction in ShortTime M odulation Domain, Speech Communication Vol 50, 2008, pp.453446. [5] Kamath S, Loizou P, A M ultiband Spectral Subtraction M ethod for Enhancing Speech Corrupted by Colored Noise, Proc.. IEEE Intr.conf. Acoustics, Speech Signal Process.vol30, 1982, pp.679681. [6] Eric Plourde, Benoit champagne, Auditory based Spectral Amplitude Estimators for Speech Enhancement, Proc.. of IEEE Trans on ASL, Vol.16 , No.8 [7] Y. Ephraim D. M alah Speech Enhancement using a M inimum M eanSquare Error spectral Amplitude estimator, Proc. of IEEE Trans on ASS, Vol. 32, No.6, Dec 1984. pp. 1109 1121. [8] Y. Ephraim D. M alah Speech Enhancement using a M inimum M eanSquare Error Logspectral Amplitude estimator, Proc. of IEEE Trans on ASS, Vol.33, No.2, April 1985, pp. 443445. [9] Chang Huai You, Soo Ngee Koh, Susanto Rahardja, Î˛ order MM SE Spectral Amplitude Estimation for Speech Enhancement, Proc.. of IEEE Trans.. on speech and Audio Processing, Vol.13, No.4 , July 2005. [10] V.Sunny Dayal, T.Kishore Kumar, Speech Enhancement using Subband wiener filter with Pitch Synchronous analysis. IEEE conference, 2013. [11] Eric Plourde, Benoit Champange, Generalized Bayesian Estimators of the spectral Amplitude for speech
I.J. Intelligent Systems and Applications, 2016, 12, 3745
A Survey on Speech Enhancement M ethodologies
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
Enhancement, IEEE signal Processing Letter, Vol 16, No 6, June 2009 Timo Gerkman, M artin Krawczyk, MM SEOptimal Spectral Amplitude Estimation Given the STFT Phase, IEEE signal Processing Letter, Vol . 20, No 2, Feb 2013 Shan An, Changchun Bao, Bingyin Xia An Adaptive βorder MM SE Estimator for speech Enhancement using SuperGaussian Speech M odel. Thomas Lotter, Speech Enhancement by MAP Spectral Amplitude Estimation Using a SuperGaussian Speech M odel, EURASIP Journal on Applied Signal Processing 2005 Vol.7, pp. 11101126 Kevin W. Wilson, Bhiksha Raj, Paris smaragdis, Ajay Divakaram. Speech Denoising Using Nonnegative M atrix Factorization with Prior, proc . of ICAASP, 2008 N. M ohammadiha, T. Gerkman, A. Leijon, “A New Linear MM SE Filter for Single Channel Speech Enhancement Based on Nonnegative M atrix Factorization,” IEEE Workshop Applications of Signal Process. 2011: 4548 N. M ohammadiha, P.Smaragdis, A. Leijon, Supervised and Unsupervised Speech Enhancement Nonnegative M atrix Factorization, IEEE Trans on Audio, Speech, and Language process, Vol. 21, No. 10 oct 2013, pp 21402151. Y. Ephraim, M urray Hill, D. M alah, On the application of Hidden M arkov M odels for enhancing Noisy Speech, Proc. of IEEE Trans . on ASS, Vol. 37, No. 12, Dec 1989, pp.18461856. Sunnydayal. V, N. Sivaprasad, T. Kishore Kumar, A Survey on Statistical Based Single Channel Speech Enhancement Techniques, IJISA, Vol 6, No.12, November 2014 Ephraim Y, M alah D., On the Application of Hidden M arkov M odels for Enhancing Noisy Speech, IEEE Trans. on ASS, 1989, Vol 37 No.12: 18461856 Balazs Fodor, Tim Fingscheidt, Speech Enhancement using a joint M AP Estimator With Gaussian M ixture M odel For NON Stationary Noise, Proc. of ICAASP 2011. Timo Gerkmann, M artin KrawczykBecker, and Jonathan Le Roux, Phase Processing for Single Channel Speech Enhancement, IEEE signal processing Magazine, Vol. 32 No.2, M arch 2015, pp. 5566 Timo Gerkman, Bayesian Estimation of Clean Speech Spectral Coefficients Given Apriori Knowledge of Phase. IEEE Trans. on Signal Processing, Vol 62, No 16. 41994226 Sunny Dayal Vanambathina, T. Kishore Kumar, Speech Enhancement using a Bayesian Estimation Given Apriori Knowledge of Clean Speech Phase. Speech com., November 2015.
45
of Engineering, Mylavaram, India. He has 4 years of teaching experience. He has 2 International publications. His interest area of research is speech processing
Dr. P. V S ubbaiah was graduated in ECE from Bangalore University and received his M aster’s Degree from Andhra University, Visakhapatnam in 1982. JNTU, Hyderabad has conferred Ph.D degree on P.V. Subbaiah for his work on M icrowave Antenna Test Facilities in the year 1996. He has vast teaching experience of 33 years in different reputed Institutions as Assistant Professor, Associate Professor, Professor and Head of the Department and Principal. Presently he is the Professor of ECE at V.R. Siddhartha Engineering College, Vijayawada and discharging his duty as the Coordinator of World Bank funded TEQIP Project since 2014. His areas of interest include M icrowave Antennas, Smart Antennas and Communications. He has published more than 100 research papers in National and International Journals and Conferences of repute. Ten research scholars have received their Ph.D degree under his supervision and presently guiding three more scholars for their Ph D. He is the M ember and Fellow of various professional societies namely ISTE, BM ESI, IETE and IE (I). He was recipient of Sir Thomas ward Gold Prize from The Institution of Engineers (India).
How to cite this paper: Ravi Kumar. K, P.V. Subbaiah, "A Survey on Speech Enhancement M ethodologies", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.3745, 2016. DOI: 10.5815/ijisa.2016.12.05
Author’s Profiles Ravi Kumar Kandagatla was born in M arkapur, India in 1988. He received the Bachelor of Technology degree from Jawaharlal Nehru Technological University, Kakinada in 2009 and received M aster of Technology in Digital Electronics and Communication Systems from Jawaharlal Nehru Technological University, Kakinada in 2011. He is presently working as Assistant professor in Lakireddy Balireddy College
Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 3745
I.J. Intelligent Systems and Applications, 2016, 12, 4656 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.06
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT Yingwei Xiao School of Electrical Engineering and Automation, Hefei University of Technology, Hefei, 230009, China Email: xiaoyingwei8@126.co m, xiaobenxian@126.co m
Abstract—Transient analysis is carried out for the singlephase grounding fault in small current grounding system, the transient grounding current expression is derived, and the influence factors are analyzed. Introduces a method for nonstationary and nonlinear signal analysis method –Hilbert Huang transform (HHT) to analyze the single phase grounding fault in small current grounding system, HHT can be better used to extract the abundant transient time frequency information fro m the nonstationary and nonlinear fau lt current signals. The empirical mode decomposition (EMD) process and the normalized Hilbert Huang transform (NHHT) algorith m are presented, NHHT is used to analyze and verify an example o f the nonlinear and nonstationary amplitude modulation signals. Bu ild a s mall current grounding system in the EMTP_ATP environ ment, by selecting the appropriate time window to ext ract the transient signals, NHHT is used to analyze the transient current signals, and the Hilbert amplitude spectrum and the Hilbert marginal spectrum of the zero sequence transient current signals are obtained. Finally, the influences of the fault phase and the grounding resistance on the time frequency characteristics of the signals are analyzed. Index Terms—Transient component of zerosequence current, fault phase, grounding resistance, NHHT, intrinsic mode function (IMF), EM D, Hilbert marg inal spectrum.
I. INT RODUCT ION At present, the most widely used signal processing method is Fourier transform, but since Fourie r transform can only be applied to linear superposition of FM signal, it can not do time frequency analysis, in order to solve this problem, wavelet transform came into being. The wavelet transform proposed by Morlet et al., through the expansion and trans lation of the signal to carry out mu ltiscale deco mposition, can effectively obtain all kinds of time frequency information fro m the signal, it has good localizat ion property in time do main and frequency domain, and has the characteristic of multiresolution analysis. However, wavelet analysis is not adaptive, once the wavelet base and the decomposition scale are chosen, it is Copyright © 2016 MECS
used to analyze the data with mult iple frequency components, the results can only reflect the signal characteristics of a fixed timefrequency resolution, and wavelet analysis is not suitable for nonstationary data. A new signal processing method  Hilbert Huang transform[1], is proposed for the nonstationary signal by Norden E. Huang et al. In 1998, this method is considered to be a major breakthrough in the linear and steadystate analysis of the Fourier transform. Norden E. Huang at first proposed the HHT to analyze ocean current signal and wave signal data, and proved its advantages in the analysis of nonstationary and nonlinear signal processing field. Since then, he has used HHT to analyze seismic wave data, wh ich greatly inspired many earthquake researchers and applied it to the field of earthquake research[23]. The application of HHT is very extensive, such as Analysis of earth's inner force wave in the field of geophysics[4], structural identification and modal analysis in structural analysis[5], the analysis of solar radiation changes in the field of astrophysics[6], the estimation of Teager energy by using the Hilbert–Huang transform[7], and in [8], HHT is used as an adaptive method to study their mult iple scale dynamics in Marine environmental time series, and so on. Its good timefrequency characteristics are demonstrated. In the biomedical field, emp irical mode decompos ition method is used to analyze the electrogastrogram in [9], emp irical mode decomposition is used to analyze heart rate variab ility & b lood pressure in [10], the HHT is used to extract new heath indicators from stationary/ nonstationary vibration signals in [11], Hilbert Huang transform (HHT) is used to improve the spectrum estimates of heart rate variability (HRV) and extract the features of HRV signals in [12]. In the field of fault processing, the applicability of Hilbert–Huang transform (HHT) for internal leakage fault detection in valvecontrolled hydraulic actuators is investigated in [13], the HHT is proposed to detect the very early stage fault in interturn insulation by only analyzing the stator current in the PMSW G in [14], fault detection is realized by means of HHT of the stator current in a PMSM with demagnetization in [15]. Also, Wavelet and Artificial Neural Networks are used to Fault diagnosis [1617]. In the field of energy, Hilbert spectral analysis (HSA) I.J. Intelligent Systems and Applications, 2016, 12, 4656
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
is used to characterize the time evolution of nonstationary power system oscillations following large perturbations in [18], also HHT is used to analyze the nonstationary powerquality waveforms in [19]. The core part of HHT algorith m is emp irical mode decomposition (EM D). In essence, EMD is to gradually decompose the oscillat ion mode or trend of d ifferent scale in signal, and produces a series of time series with different characteristic scales, and each t ime is an IMF component of intrinsic mode function. With the help of Hilbert transform (HT), the time spectrum of signal can be further obtained, which can accurately reflect the timefrequency characteristics of the signal. In the system of 3 ~ 66kV power system in wh ich the neutral point is not connected to ground or connected to ground through the arc suppression coil, when the ground fault occurs in a certain phase, because the short circuit can not be formed, the fault current is often much smaller than the load current, this phenomenon is called "small current grounding system". The practice shows that the occurrence of singlephase grounding fault in the power grid is the highest, which is about 85% of the total failure of the system. Because the fault current is small in the s mall current grounding power system, and the detection is difficult, and has not been well solved, which has become one of the research hotspots in this field. In this paper, HHT is used to analyze the fault current of singlephase ground fault in small current grounding system. The main research work of this paper is as follows: (1) The transient analysis of the grounding current is carried out and the influence factors are analy zed, the grounding current is main ly transient capacitance current and transient inductance current, but because of the difference between the two frequencies, it can not be offset each other. (2) The emp irical mode decomposition (EMD) process and the normalized Hilbert transform (NHT) algorith m are presented. The NHHT is used to analyze and verify the nonlinear and nonstationary amplitude modulation signals. (3) Under the EM TP_ATP environ ment, a small current grounding system is built, and the parameters of the system are calcu lated, and the singlephase grounding fault is simulated. (4) Extracts transient signal, use HHT to analy ze transient current signal, and obtain time spectrum and Hilbert marginal spectrum of Zero sequence transient current signal of fault feeder, and the influence of the fault phase and the grounding resistance on the timefrequency characteristics of the signal is analyzed.
II. TRANSIENT ANALYSIS OF SINGLE PHASE GROUNDING FAULT IN SMALL CURRENT GROUNDING SYST EM When single phase grounding fault occurs in the neutral point via the arc suppression coil grounding grid, transient capacitance current and transient inductance current constitute the transient grounding current flowing Copyright © 2016 MECS
47
through the fault point. But because of the difference between the two frequencies, the transient process can not compensate for each other. The equivalent circuit of system transient current for the neutral point via the arc suppression coil grounding grid is shown in Fig.1. L0
R0
RL
~
iC
u0 iL
C0
L
Fig.1. T ransient equivalent circuit of single phase grounding fault
In Fig.1, C0 is the equivalent capacitance of the grid to the ground, L0 is the equivalent inductance of the circuit, power supply, and the transformer in the zero sequence circuit. R0 is the equivalent resistance of the zero sequence circuit, including the grounding resistance of the fault point and the arc path resistance. RL and L are respectively the active loss resistance and inductance of the arc suppression coil, u0 is zero sequence power supply voltage. A. Transient capacitance current Since the self vibration frequency of the capacitance current is generally h igher, and the inductance of the arc suppression coil is L L0 , so the influence of RL and L in Fig.1 is neglected in analy zing its transient characteristics. By using the equivalent circuit of L0 &
C0 & R0 and the system zero sequence voltage source u0 U m sin(t ) , transient capacitance current iC can be obtained. From Figure 1, the differential equation can be got:
R0iC L0
diC 1 dt C0
t
i d u 0 C
(1)
0
When the grounding resistance is very small, at R0 2 L0 C0 , the system is in a state of under damping condition, the transient capacitance current has a periodic oscillation and attenuation characteristics. Transient self vibration component iC.os and steady state power frequency
component
iC.st
constitute
transient
capacitance current iC , by using the relationship between iC.os iC.st 0 and I Cm U mC0 at init ial conditions t 0 , the follo wing equation can be got by Laplasse transformation: iC iC.os iC.st I Cm f sin sin t cos cos f t e t cos(t )
(2)
I.J. Intelligent Systems and Applications, 2016, 12, 4656
48
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
In the formu la (2): U m is the amp litude of the phase voltage, I Cm is the amplitude of the capacitor current, f is the transient selfvibration angular frequency, 1 C R0 L0 2 is the attenuation coefficient of free oscillating co mponent, where the C is the time constant of the loop. When C is larger, the self  vibrat ion attenuation is slower, whereas, the attenuation is faster. By analy zing formula (2), if the single phase grounding fault occurs in 2 , and at t T f 4 ( T f 2 f is self vibrat ion period), the amplitude of the self vibrat ion part of the capacitive current is the maximu m
f 4T e
iC.os max :
L st
L Z
t
[cos( )e
L
cos(t )]
(7)
In the formula (6): st U m (W ) is the steady state flu x, arctan[ RL (W )] is the phase angle of compensating
current, Z RL2 ( L)2
suppression coil impedance,
is
the
arc
L is the time constant of
the inductance loop. Due to RL L , take Z L , 0 . In consideration of L os st , iL iL dc iL st and I Lm U m ( L) , the exp ression of the transient inductance current iL can be obtained as follow:
Tf
iC.os max I Cm
(3)
C
If the single phase grounding fault occurs in 0 , and at t T f 2 , the min imu m value of iC.os min for the self vibration current is:
f 2T e Tf
iC.os min I Cm
(4)
C
Selfvibrat ion frequency f of selfvibrat ion current is:
f 02 2
R 1 0 L0C0 2 L0
2
(5)
In formu la (2), Selfvib ration frequency of the loop is 0 1 L0 C0 . When the grounding resistance increases to R0 2 L0 C0 , the system is in a state of over damping state, and the loop current has a non periodic oscillat ion attenuation characteristic and tends to be stable. According to the above analysis, the change of the fault resistance has a great influence on the transient characteristics of the grounding system of the neutral point via arc suppression coil.
t
iL I Lm [cos e L cos(t )]
(8)
By the formula (8), it is known that the transient DC component and the steadystate AC co mponent constitute the transient inductance current iL , and the initial amp litude of the transient DC decay process is related to the phase , the attenuation coefficient is determined by loop impedance. Transient DC co mponent in the 0 is the largest, and in the 2 is the smallest. C. Transient grounding current System transient grounding current is composed of transient capacitance current iC and transient inductance current iL , although the amplitude difference between iC and iL is not large, they are in different frequency bands and can not compensate for each other. The expression of the transient grounding current id can be derived from the formula (2) and (8):
id iC iL ( I Cm I Lm ) cos(t ) f I Cm sin sin t cos cos f t e C I Lm cos e L (9) t
t
B. Transient inductance current Differential equation of the core flu x can be derived from Fig. 1 [1]: u0 RLiL W
d L dt
(6)
In the formula (6), W is the turns of the arc suppression coil, and L is the core flu x of the arc suppression coil. Before the grounding fault, no current flo ws through the arc suppression coil, that is, L is zero. Take iL W L L into formu la (6), L expression can be obtained as follow:
Copyright © 2016 MECS
In the formula (9), the first part of equality right is the steadystate component of the grounding current, the amp litude is the difference between the amplitude of the steady state capacitance current and the steady state inductance current, and is closely related to the fault phase , frequency is power frequency. The other part is the transient component of the grounding current, wh ich is composed of two parts, transient free oscillat ion component of capacitance current and DC decaying component of inductance current. Fro m the above analysis, after the occurrence of a single phase grounding fault, transient grounding current flowing through the fault point contains transient capacitive current component of damped oscillation and the decay transient inductance current co mponent. Whether the system is operated in the neutral point via I.J. Intelligent Systems and Applications, 2016, 12, 4656
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
49
the arc suppression coil grounding mode or not grounding mode, the transient capacitance current dominates the amp litude and frequency of the transient grounding current, and its amplitude is also related to the fau lt phase. The init ial value of DC co mponent in the transient inductance current is related to the initial phase and core saturation.
envelope which are determined by local maximu m point and local minimu m point is zero, that is, the signal is locally symmetric about the time axis. IMF is essentially a narrow band signal. Because any given signal can not satisfy the IMF condition, the comp lex signal is deco mposed into a series of IMF by EM D[20]. For a g iven signal X (t ) , the EMD process is as follows[21]:
III. HILBERT HUANG T RANSFORM (HHT)
Step1. finds out all the local maxima and minima of X (t ) , by using the three spline interpolation, the local extreme value is connected to the upper and lower two envelope of emin (t ) and emax (t ) . X (t ) minus the mean m1 (t ) of the envelope, m1 (t ) (emin (t ) emax (t )) 2 , and
A. Instantaneous frequency and intrinsic mode function (IMF) Instantaneous frequency is one of the most important concepts in time frequency analysis. The defin ition of instantaneous frequency is needed to satisfy a certain condition, and not all the instantaneous frequency of the signal has a physical meaning, only when the signal is approximated by a single component signal, each mo ment corresponds to a single frequency component, the instantaneous frequency has the actual physical mean ing. In order to obtain the instantaneous frequency of physical meaning, Norden.E.Huang et al. proposed the concept of intrinsic mode function (IMF). Using Hilbert transform to analyze the intrinsic mode function (IMF) can get good time frequency characteristics, but most of the signal is not the intrinsic mode function (IMF), and is the co mbination of the many of intrinsic mode function (IM F). Therefore, the correct time frequency characteristics can not be obtained directly through the Hilbert transform, so that we can not accurately analyze the signal directly. In order to use the Hilbert transform to analyze the nonlinear and nonstationary signals correctly, the signal is decomposed to IMF at first, then the emp irical mode decomposition (EMD) is developed. B. The principle of empirical mode decomposition (EMD) Emp irical mode deco mposition (EMD) uses selection after layer by layer, and get a series of intrinsic mode functions (IMF), and the dominant frequency of the intrinsic mode function (IMF) is lower and lo wer, the dominant frequency of IMF obtained firstly is the highest, and the IMF obtained lastly is the lo west, therefore, it is essentially a filter bank. If a signal with random noise is decomposed by the emp irical mode decomposition (EMD), the high frequency intrinsic mode function (IMF) co mponent is usually a high frequency noise of the signal, and the residual signal co mponent is usually the mean or trend of the original signal. Each intrinsic mode function (IM F) component obtained from deco mposition has obvious physical mean ing and contains a certain range of characteristic scales. The intrinsic mode function (IM F) must meet the signal of the following two conditions [1]: (1) The zero pole po ints in the signal are equal or have difference at most 1. (2) Any point on the signal, the mean value of the Copyright © 2016 MECS
the first component h1 (t ) X (t ) m1 (t ) is obtained. Step2. check whether h1 (t ) is IMF. If not, then return to step 1, and take h1 (t ) as a new signal for screening. Repeat screening k times, until h1k (t ) becomes IMF, take c1 (t ) h1k (t ) . Residual signal is r1 (t ) X (t ) c1 (t ) .
Step3. take r1 (t ) as a new signal, re perform step 1 to 2, get the new IMF c2 (t ) and the residual volu me r2 (t ) . Repeat iteration n t imes: rn (t ) rn1 (t ) cn (t ) , until rn (t ) is a monotonic function, EMD is over. EM D is an adaptive filtering process, and the signal is decomposed adaptively according to the frequency scale, a series of narrow band signals are obtained fro m large to small frequency. Due to the co mpleteness of EMD, the original signal X (t ) can be expressed as the sum of n IMF and an average trend component rn (t ) . C. The normalized Hilbert Huang transform (NHHT) The normalized Hilbert  Huang transform (NHHT) is an improved Hilbert Huang transform (HHT), it is composed of empirical mode deco mposition (EM D) and normalized Hilbert transform (NHT). For any continuous time signal X (t ) , Hilbert transform is as follow: Y (t )
1
X ( ) d t
(10)
X (t ) and Y (t ) form co mplex conjugate pairs, and constitute an analytical signal as follows: Z (t ) X (t ) jY (t ) X 2 (t ) Y 2 (t )e j (t ) (t ) arctan
Y (t ) X (t )
(11) (12)
According to the derivative defin ition of the analytic signal phase, the instantaneous frequency is as follow:
I.J. Intelligent Systems and Applications, 2016, 12, 4656
50
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
f (t )
1 d (t ) 2 dt
(13)
It can be seen that f (t ) is a single valued function of t , the frequency changes with time, in order to make the instantaneous frequency meaningfu l, the time series of the Hilbert t ransform must be monotonous, and the IMF after the EMD just meets the requirement. The definition of Hilbert spectrum (united timefrequency amplitude spectrum)[22] is: k
k
H ( ,t)= Re zi (t)=Re ai (t )e ji (t ) i =1
(14)
i =1
Hilbert marginal spectrum[23] is: T
H ( )= H (,t)dt
(15)
0
In the formu la (15), T is the signal length. Based on the H (,t) energy distribution in time frequency space, the energy (amp litude) of all mo ments for a certain frequency is added to form the total energy (amp litude) of the signal, that is, the Hilbert marginal spectrum H ( ) . But the frequency is not always present at all time, or only at a certain mo ment, it may appear several t imes in d ifferent or equal amp litude, this physical mean ing is consistent with the physical mean ing of the Fourier spectrum[24], but in the Fourier spectrum, the same magnitude is required for any frequency, which is easy to destroy the true frequency of the signal and appear false frequency. Because of the limitation of Bedrosian principle , the amp litude can cause serious disturbance to frequency modulation, wh ich affects the instantaneous frequency. In order to eliminate this limitation, a normalized algorith m is adopted to decompose IMF into a unique empirical envelope (AM) and a carrier (FM), and then Hilbert transform for carrier is done for Instantaneous frequency[25]. The essence is a kind of reverse frequency modulation and amplitude modulation. Specific steps are as follows: Step1. The maximu m value of the absolute value of each IMF co mponent is solved, and the endpoint is also regarded as an extreme value. Step2. The three spline is used to fit the maximu m value curve e1 (t ) . When the envelope is lower than the original data sequence, the local area envelope is rep laced by a straight line envelope, use envelope normalized data f1 imf e1 , f1 is used as the new IMF. Step3. Repeat steps 1, 2 until the absolute value of IMF is no more than 1, and eventually will obtain its frequency modulation (carrier) co mponent f n and amplitude modulation components a(t ) e1 e2 en . Step4. In order to obtain the instantaneous frequency, the Hilbert transform is carried out for the carrier. United timefrequency amplitude spectrum of the Copyright © 2016 MECS
signal is obtained basedon amplitude modulation component and instantaneous frequency, and then the Hilbert marginal spectrum is obtain. D. Analysis of an example As a time frequency analysis method, of course, NHHT is most commonly used in nonlinear non stationary signal analysis. Use NHHT to analyze nonlinear and nonstationary AM and FM signal in the formula (16): Y sin(30 t ) 10t cos(70 t sin(20 t )) sin(30 t 2 ) (16)
By the formu la (16), it is known that the signal is composed of a 15Hz sinusoidal signal, a linear FM signal and an AM & FM signal. Through the FFT analysis, the Fourier spectrum (Fig.2) was obtained, and the amplitude peak values were appeared at 15Hz, 25Hz, 35Hz, 45Hz, 55Hz frequency in Fourier spectrum, and the amplitude of 35Hz was the largest.
Fig.2. Fourier spectrum of the simulation signal
Use NHHT to analy ze the signal, firstly EM D (Fig. 3) is done, the original signal is decomposed into a series of IMF and residual RES, IM F is a s eries of signal components with different frequency components. According to the approximate orthogonality of EMD, the IMF is nearly orthogonal to each other, that is, the frequency components of each IMF co mponent are almost different. In Fig.3, each IM F is arranged according to its frequency components fro m the big to small arrangement, and due to the residual RES is very small, the decomposition is more successful. Seen fro m Fig.3, the signal is made up of three parts: the 15Hz constant frequency component, linear frequency modulation component, as well as frequency oscillation co mponent by taking 35Hz as the center and 10Hz as the oscillat ion amplitude. NHT is done for the IMF obtained by EMD, and united timefrequency amp litude spectrum (Fig.4) is obtained. In the united time frequency amplitude spectrum, the amp litude changes of the same frequency can be seen at different times, this is not done in the Fourier analysis which is required to be based on the same frequency and its amplitude can not be changed with time. Seen fro m Hilbert marginal spectrum (Fig.5), there is no mo re than 50Hz frequency band, and the frequency components of 55Hz appear in the Fourier spectrum. According to the theoretical analysis, in a second the signal frequency can not be greater than 50Hz, the false I.J. Intelligent Systems and Applications, 2016, 12, 4656
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
frequency appears in the Fourier spectrum, actual results show that the FFT is not suitable for analy zing nonlinear signals and appear spectrum leakage, and NHHT is not restricted, these show, compared to FFT and other conventional algorith ms, NHHT in analyzing nonlinear signal has incomparable superiority.
51
ATP is the program package programmed by the Danish scholar Hart basedon the use of EMTP for the kernel. Because ATP uses the Windows interface, it has a more friendly manmachine window co mpared with the traditional EMTP, the application is more simple. In ATP, the simu lation of 5 output lines is built, and the topological structure of the model is shown in Fig.6. Fig.6 simu lation model is a 10kV threephase symmetric system model, the line uses the distribution parameters module with 5 outlets. In Fig.6, 10kV voltage is obtained by 110kV A C 3 phase power supply through the 110/10kV transformer, the transformer is Yy 0. It is not the grounding system when the K is open, and is the grounding system through the arc suppression coil when the K is closed. T
Line1=15km Line2=6+18km
110KV
10KV
Line3=30km
K
Fig.3. EMD of the simulation signal
Line4=18km
RL L
Line5=43km Rf
Fig.6. Simulation model of small current grounding fault
Fig.4. United timefrequency amplitude spectrum of the simulation signal
In Fig.6, the parameters of the s mall current fault grounding system are shown as fo llo ws: Line L1 is 15km pure cable line. Line L2 is a mixed line of overhead line and cable, the length of which is 18km and 6km respectively. L3, L4 and L5 are the pure overhead lines, the length of which is 30km, 18km and 43km. Cable line parameters are: Positive sequence impedance Z1 (0.125 j 0.095) / km , Zero sequence impedance Z0 (0.97 j1.590) / km . Positive sequence susceptance b1 83.566 s / km , Zero sequence susceptance b0 83.566 s / km . The parameters of the overhead line are: Positive sequence impedance Z1 (0.270 j 0.351) / km ,
Fig.5. Hilbert marginal spectrum of the simulation signal
Zero sequence impedance Z0 (0.475 j1.757) / km . Positive sequence susceptance b1 3.267 s / km , Zero
IV. SIMULAT ION AND A NALYSIS OF SINGLE PHASE GROUNDING FAULT
sequence susceptance b0 1.100 s / km . The terminal load of each outlet is 1000 j 20 . The capacitance current of single phase to ground can be calculated as follow:
A. Construction of single phase grounding fault system EMTP is the first of the electromagnetic transient analysis software produced by Tan Weier, Canada, wh ich is a powerful tool for power system simu lation, wh ich can be used to analyze the steady and transient characteristics of power system. The typical application of EMTP is to simulate the change of electrical parameters with t ime under the disturbance in power system, such as the change of the feeder current after the occurrence of grounding fault in the power grid. Copyright © 2016 MECS
I c 3C0U 32.46(A) 20(A)
(17)
In general, when the capacitance current is more than 20A in the 10KV mediu m voltage power grid, it is necessary to install the arc suppression coil. The inductance of the arc suppression coil is obtained by using formula (18):
% L
1 3C0
(18)
I.J. Intelligent Systems and Applications, 2016, 12, 4656
52
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
The arc suppression coil is set by the over compensation of % =110%, through calculation, the inductance of arc suppression coil is L =0.55H. The value of series resistance is set to RL =17.28 by the 10% arc
steady state and transient components. Can be seen after the above treatment, it is very good to filter out the steadystate component of the fault current, the transient component is obtained as shown in Fig.7 (B).
suppression coil inductance.
2) Transient time frequency analysis of zero sequence transient current
B. Time frequency analysis of zero sequence transient current based on HHT 1) Extraction of transient signal Fro m the previous analysis, the transient grounding current includes the steady state power frequency component and transient component of grounding current. According to the superposition principle, the power frequency component in the fault transient zero sequence current of each feeder is the superimposition of the asymmetric co mponent before the fault and the steady state power frequency component after the fault. And because the actual small current grounding fault occurs after the 5~6 cycle, as shown in Fig.7 (a), the zero sequence current transient component has been quite small, it is considered that the electromagnetic transient process is basically the end. Therefore, a reasonable time window is designed to filter the steadystate component, its window size is a power frequency cycle, the expression of the zero sequence current transient component is as follow:
i0 kf (t ) i0 k (t ) i0 k (t 7T )
(19)
Zero sequence current/A
In the formula (19), i0 k (t ) is the zero sequence current of the k feeder, and T is the power frequency cycle.
Time/s Zero sequence transient current/A
(a) T he zero sequence current of fault line
Time/s
(b) T he Zero sequence transient current of fault line Fig.7. T he extraction of fault signal transient component
In Fig.7 (a), the grounding fault occurs in the feeder L5 at the bus terminal 12km and the fault phase 600 , the ground resistance is 2 . The actual zero sequence current flows through the feeder L5, wh ich contains the Copyright © 2016 MECS
HHT is done for the zero sequence transient current signal in Fig.7 (b). First, EMD is done for the zero sequence transient current, seen from Fig.8 (a), the current signal which contains a variety of frequency is decomposed into a series of different center frequency current components, namely IMF1 to IM F6, its time scale is getting bigger and bigger, that is, the center frequency of each IMF frequency band is gradually reduced. The center frequency and bandwidth of each current component are adaptive, without a priori decomposition basis function, the frequency components of each IMF component are not equal in theory, and the amp litude of each IMF component is determined by its time and frequency. From the ratio of the EM D residual amplitude and the original signal, the deco mposition is more successful. Fro m Fig.8 (a), the main frequency components of the signal are concentrated in IMF4, the amp litude of the other IMF co mponents are very s mall which co mpared to IMF4, also the frequency is about 190Hz by Hilbert transform for IM4, wh ich is relative to the Hilbert marginal spectrum of Fig.8 (c). The united timefrequency amplitude spectrum is obtained by the normalized Hilbert transform (NHT) for the IMF co mponent of the decomposition, that is, the Hilbert spectrum (Fig.8 (b)). Fro m Hilbert spectrum, it can be seen that the distribution of the frequency components of the transient current in the whole t ime frequency space, the main frequency band of the transient current is concentrated in the low frequency band between 0 and 500Hz. The h igh frequency components of the frequency components are relatively scattered and mainly concentrated in the mo ment of failu re occurrence. In the Hilbert spectrum, the amplitude is s maller but the power frequency steadystate periodic co mponent is found in the whole analysis. This corresponds to IMF6 after EM D. Fro m Fig.8(a), it can be seen the frequency of IMF6 is 50Hz, and the analysis of the transient current in the last part shows that this is the function of the system inductance. The Hilbert spectrum shows that the signal is attenuated in the whole, and the performance of the first 1/4 power frequency cycle is obvious, which is consistent with the transient component of zero sequence current. The amp litude spectral of Hilbert is integrated in the whole analysis, and the amp litude of each frequency component at each mo ment is accu mulated, the Hilbert marginal spectrum is obtained. Th is is a kind o f spectrum of signal energy d istribution, it is a kind of characterization form o f frequency and energy, and the amp litude will be normalized in the practical application. Seen fro m Fig.8 (c) o f the Hilbert marg inal spectrum, its frequency components are relatively concentrated around 200Hz, the main energy d istribution of the signal is between 0500Hz. I.J. Intelligent Systems and Applications, 2016, 12, 4656
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
53
is changed at fault line (L5). Fro m the graph, we can know that with the increase of the fault phase, the DC component of the current decreases rapidly, and the amp litude of the current is obviously increased, which is consistent with the conclusion of the theoretical analysis. It can be seen fro m the Hilbert spectrum of different fault phase, the amplitude of the main vib ration mode of the current signal is exponential decay with the increase of time. W ith the increase of the phase, the frequency components of the current are beco ming more and more abundant, and the marginal spectrum of Hilbert describes this phenomenon in the frequency domain.
i/A
(a) Empirical mode decomposition of zero sequence transient current
(b) United timefrequency amplitude spectrum of zero sequence transient current
t/s
f/Hz
i/A
(a) Hilbert amplitude spectrum at 00
f/Hz
(c) Hilbert marginal spectrum of zero sequence transient current
(b) Hilbert marginal spectrum at 00
C. The influence of main factors on the transient timefrequency characteristics
i/A
Fig.8. T ransient time frequency analysis of fault line (L5) zero sequence current
The main factors that affect the fault signal are the structure of the circuit, the fault phase, the grounding resistance, the distance between the fault point and the bus, among them, the fault phase and grounding resistance are the dominant factors. In this paper, the effects of the fault phase and the grounding resistance on the timefrequency characteristics of the signal are discussed.
Under the condition of grounding resistance 5, by changing the fault phase, the zero sequence current of a series of fault feeder (L5) is obtained. The transient component is obtained by the previously designed filter window, the timefrequency characteristics of the current signal are obtained by HHT analysis, as shown in Fig.9. Fig. 9 is time frequency characteristics of the transient component of zero sequence current when the fault phase Copyright ÂŠ 2016 MECS
i/A
1) The influence of the fault phase on the transient timefrequency characteristics
t/s
f/Hz
(c) Hilbert amplitude spectrum at 300
f/Hz
(d) Hilbert marginal spectrum at 300
I.J. Intelligent Systems and Applications, 2016, 12, 4656
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
i/A
54
f/Hz
t/s
i/A
(e) Hilbert amplitude spectrum at 600
Fig.10 is Hilbert spectrum and Hilbert marg inal spectrum of the fault line at 50 , 100 , 500 , 5000 . It can be seen from the figure that with the increase of grounding resistance, signal current amp litude and frequency components gradually reduced. At the 50 around, the system has been in a state of damping, and the frequency components of the system are re latively concentrated in Fig.10 (b). And with the increase of the grounding resistance, the sys tem is re entered into an under damping condition, as shown in Fig.10 (d), although the main frequency components of the system are concentrated in 50Hz, but the energy of 150Hz and 250Hz is also larger, and the system energy is relatively dispersed. After a further increase in the grounding resistance, the current signal frequency is further concentrated in the low frequency part, as shown in Fig.10 (f), and the signal energy becomes more concentrated. When the grounding resistance reaches 5000 values, it almost beco mes a DC co mponent, as shown in Fig.10 (h).
f/Hz
i/A
i/A
(f) Hilbert marginal spectrum at 600
f/Hz
t/s
i/A
i/A
t/s
f/Hz
(a) Hilbert amplitude spectrum at 50
(g) Hilbert amplitude spectrum at 900
f/Hz f/Hz
(h) Hilbert marginal spectrum at 900
(b) Hilbert marginal spectrum at 50
In short, with the increase of the phase, the amplitude and frequency components of the signal increase, the distribution of the signal energy will be more and more dispersed.
i/A
Fig.9. T he influence of the fault phase on the transient timefrequency characteristics of fault line zero sequence current
2) The influence of the grounding resistance on the transient timefrequency characteristics f/Hz
In [26], the effects of veryhigh resistance grounding on the selectivity of groundfault relaying are studied. The fault resistance and fault phase are the main factors that affect the zero sequence current. Set the fault phase at 900 , change the grounding resistance, analyze the signal timefrequency characteristics of different grounding resistance. Copyright © 2016 MECS
t/s
(c) Hilbert amplitude spectrum at 100
I.J. Intelligent Systems and Applications, 2016, 12, 4656
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
55
i/A
towards the low frequency in general, the frequency components are more and more concentrated, and the energy distribution is beco ming more and more concentrated.
V. CONCLUSION
f/Hz
i/A
(d) Hilbert marginal spectrum at 100
t/s
f/Hz
i/A
(e) Hilbert amplitude spectrum at 500
f/Hz
i/A
(f) Hilbert marginal spectrum at 500
t/s
f/Hz
(g) Hilbert amplitude spectrum at 5000
In this paper, the transient analysis of singlephase grounding current in small current grounding power grid is carried out. Resultly the grounding current is mainly capacitive current and inductance current. On this basis, the theoretical analysis and simulat ion are done. According to the different fault condit ions, such as feeder structure, fault phase, grounding resistance, fault distance, etc, the simulation results show that the fault phase and the grounding resistance have most influence on the feeder zero sequence current. The main results of this paper are as follows: (1) Analy ze the steadystate and transient characteristics of single phase grounding fault in the small current grounding power grid, in wh ich the neu tral point is not connected to ground or connected to ground through the arc suppression coil. (2) Under the EM TP_ATP environ ment, a small current grounding system is built, and the simulat ion analysis of the singlephase grounding fault under different conditions is done. According to the power system related theory calculat ion result, the system grounding capacitance over 10KV voltage level is allo wed the maximu m capacitance current 20A, the neutral point must be connected with the arc suppression coil, and the compensation degree is 10%. (3) HHT is applied to the single phase grounding fault detection through the arc suppression coil. Firstly, a time window is selected to extract the transient component of the fault signal, and the transient component of the feeder zero sequence current is extracted, and the time frequency analysis of transient signal is done with the HHT. The two main factors that influence the transient signal, the fault phase and the grounding resistance, are studied respectively. A CKNOWLEDGMENT
i/A
The authors wish to thank my graduate tutor Hongbin Wu Professor and Benxian Xiao Professor and graduate student Lin Zhang. REFERENCES [1] f/Hz
(h) Hilbert marginal spectrum at 5000 Fig.10. T he influence of the grounding resistance on the transient timefrequency characteristics of fault line zero sequence current
In short, with the continuous increase of the grounding resistance, the system undergoes a sate of less damping, over damping, and then to less damping, the signal is Copyright © 2016 MECS
[2]
[3]
N.E.Huang, Z.Shen, S.R.Long, et al., ―The empirical mode decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis,‖ Proc R Soc Lond A, 454, pp.903995, 1998. Vasudevan K, Cook FA., ―Empirical mode skeletonization of deep crustal seismic data: Theory and applications,‖ Journal of Geophysical ResearchSolid Earth, Vol.105(B4), APR 10, pp.78457856, 2000. Loh CH, Wu TC, Huang NE., ―Application of the empirical mode decompositionHilbert spectrum method to identify nearfault groundmotion characteristics and
I.J. Intelligent Systems and Applications, 2016, 12, 4656
56
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT
structural responses,‖ Bulletin of the Seismological Society of America, Vol.91, No.5, pp.13391357, 2001. Xun Zhu, Zheng Shen, Stephen D.E, et al., ―Gravity wave characteristics in the middle atmosphere derived from the empirical mode decomposition method,‖ Journal of Geophysical Research, Vol.102, No.D14, pp.1654516561, 1997. Yang Jann N, L.Y., ―System identification of linear structures using Hilbert transform and Empirical M ode Decomposition,‖ Proceedings of the International Modal Analysis Conference, pp.213219, 2000. Komm RW, Hill F. Howe R., ―Empirical mode decomposition and Hilbert analysis applied to rotation residuals of the solar convection zone,‖ Astrophysical Journal, Vol.558, No.1, pp.428441, 2001. Ranjit A. Thuraisingham, ―Estimation of Teager energy using the Hilbert–Huang transform,‖ IET Signal Process, Vol.9, No.1, pp.82–87, 2015. Dhouha Kbaier Ben Ismail, Pascal Lazure, Ingrid Puillat, ―Advanced Spectral Analysis and Cross Correlation Based on the Empirical M ode Decomposition: Application to the Environmental Time Series,‖ IEEE Geoscience and Remote Sensing Letters, Vol.12, No.9, pp.19681972, 2015. Liang H, Lin Z, M cCallum RW., ―Artifact reduction in electrogastrogram based on empirical mode decomposition method,‖ Medical& Biological Engineering & Computing, Vol.38, No.1, pp.3541, 2000. Echeverria J C,Crowe J A,Woolfson M S,HayesGill B R., ―Application of empirical mode decomposition to heart rate variability analysis,‖ Medical& Biological Engineering& Computing, 39(2): 471479, 2001. Abdenour Soualhi, Kamal M edjaher, Noureddine Zerhouni, ―Bearing Health M onitoring Based on Hilbert– Huang Transform, Support Vector M achine, and Regression,‖ IEEE Transactions on Instrumentation and Measurement, Vol.64, No.1, pp.5262, 2015. Helong Li, Sam Kwong, Lihua Yang, Daren Huang, Dongping Xiao, ―HilbertHuang Transform for Analysis of Heart Rate Variability in Cardiac Health,‖ IEEE/ACM Transactions on Computational Biology and Bioinformatics, Vol.8, No.6, pp.15571567, 2011. Amin Yazdanpanah Goharrizi, Nariman Sepehri, ―Internal Leakage Detection in Hydraulic Actuators Using Empirical M ode Decomposition and Hilbert Spectrum,‖ IEEE Transactions on Instrumentation and Measurement, Vol.61, No.2, pp.368378, 2012. Chao Wang, Xiao Liu, Zhe Chen, ―Incipient Stator Insulation Fault Detection of Permanent M agnet Synchronous Wind Generators Based on Hilbert–Huang Transformation,‖ IEEE Transactions on Magnetics, Vol.50, No.11, pp.14, 2014. Antonio Garcia Espinosa, Javier A. Rosero, Jordi Cusid´o, Luis Romeral, Juan Antonio Ortega, ―Fault Detection by M eans of Hilbert–Huang Transform of the Stator Current in a PM SM With Demagnetization,‖ IEEE Transactions on Energy Conversion, Vol.25, No.2, pp.312318, 2010. Shu Hongchun, Tian Xincui, Dai Yuetao, ―The Identification of Internal and External Faults for ±800kV UHVDC Transmission Line Using Wavelet based M ultiResolution Analysis,‖ I.J. Intelligent Systems and Applications, 3(03): 4753, 2011. Ashwani Kumar Narula, Amar Partap Singh, ―Fault Diagnosis of M ixedSignal Analog Circuit using Artificial Neural Networks,‖ I.J. Intelligent Systems and Applications, 7(07): 1117, 2015. A. R. M essina, Vijay Vittal, ―Nonlinear, NonStationary
Copyright © 2016 MECS
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
Analysis of Interarea Oscillations via Hilbert Spectral Analysis,‖ IEEE Transactions on Power Systems, Vol.21, No.3, pp.12341241, 2006. M ohammad Jasa Afroni, Danny Sutanto, David Stirling, ―Analysis of Nonstationary PowerQuality Waveforms Using Iterative Hilbert Huang Transform and SAX Algorithm,‖ IEEE Transactions on Power Delivery, Vol.28, No.4, pp.21342144, 2013. Patrick Flandrin, ―Empirical M ode Decomposition as a Filter Bank,‖ IEEE Signal Processing Letters, Vol.11, No.2, pp.431435, 2004. Gabriel Rilling, ―On the Influence of Sampling on the Empirical M ode Decomposition,‖ IEEE Signal Processing Letters, Vol.17, No.4, pp.214218, 2005. Xiaonan Hui, Shilie Zheng, Jinhai Zhou, Hao Chi, Xiaofeng Jin, Xianmin Zhang, ―Hilbert–Huang Transform TimeFrequency Analysis in φOTDR Distributed Sensor,‖ IEEE Photonics Technology Letters, Vol.26, No.23, pp.24032406, 2014. Yingjun Yuan, Zhitao Huang, Hao Wu, Xiang Wang, ―Specific emitter identification based on Hilbert–Huang transformbased time–frequency–energy distribution features,‖ IET Commun., Vol.8, No.13, pp.24042412, 2014. Zhaohua Wu, Norden E.Huang, ―Ensemble Empirical M ode Decomposition: a noise assisted data analysis method,‖ Advances in Adaptive Data Analysis, 1(1):141, 2009. Norden E. Huang, Zhaohua Wu, ―On Instantaneous Frequency,‖ Advances in Adaptive Data Analysis, Vol.1, No.2, pp.177229, 2009. Thomas N., ―The effects of veryhigh resistance grounding on the selectivity of groundfault relaying in highvoltage long wall power systems,‖ IEEE Trans Ind Appl., Vol.37, No.2, pp.398406, 2001.
Authors’ Profiles Yingwei Xiao was born in Hefei, China in 1993. She has completed the B.Sc degree in electrical engineering and automation from Anhui Polytechnic University, Wuhu, China in 2015. Currently she is still pursuing her master degree in Hefei University of Technology, Hefei, China. Her research interests include automation of electric power system, micro grid power flow calculation.
How to cite this paper: Yingwei Xiao, "Characteristic Research of SinglePhase Grounding Fault in Small Current Grounding System basedon NHHT", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.4656, 2016. DOI: 10.5815/ijisa.2016.12.06
I.J. Intelligent Systems and Applications, 2016, 12, 4656
I.J. Intelligent Systems and Applications, 2016, 12, 5764 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.07
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems Zhengbing Hu School of Educational Information Technology, Central China Normal University, No. 152 Louyu Road, 430079, Wuhan, China Email: hzb@mail.ccnu.edu.cn
Vadym Mukhin, Yaroslav Kornaga and Yaroslav Lavrenko National Technical University of Ukraine “Kiev Polytechnic Institute”, Kiev, 03057, Ukraine Email: {v.mukhin, y.kornaga, y.lavrenko}@kpi.ua
Oleg Barabash and Oksana Herasymenko Taras Shevchenko National University of Kiev, Volodymyrska Street, 64/13, 01601, Kiev, Ukraine Email: oksgerasymenko@gmail.com
Abstract—The article deals with the issues of the security of distributed and scalable computer systems based on the riskbased approach. The main existing methods for predicting the consequences of the dangerous actions of the intrusion agents are described. There is shown a generalized structural scheme of job manager in the context of a riskbased approach. Suggested analytical assessments for the security risk level in the distributed computer systems allow performing the critical t ime values forecast for the situation analysis and decision making for the current configuration o f a distributed computer system. These assessments are based on the number of used nodes and data links channels, the number of act ive security and monitoring mechanisms at the current period, as well as on the intensity of the security threats realization and on the activation intensity of the intrusion prevention mechanis ms. The proposed comprehensive analytical risks assessments allow analyzing the dynamics of intrusions processes, the dynamics of the security level recovery and the corresponding dynamics of the risks level in the distributed computer system. Index Terms—Distributed computer systems, Security analysis, Riskbased approach.
NOMENCLAT URE α – the average intensity of the security threats for the distributed computer systems resources x1 and x2; β – the average intensity of the security mechanis ms from the group x3 and x4 activation; R0 – the initial security risk; R(t) – security risks in the distributed computer systems, changing at time t; x1 – the number of nodes in the distributed computer system; x2 – the number of data link channels; Copyright © 2016 MECS
x3 – the number of the active security mechanisms; x4 – the number of the security monitoring tools in the distributed computer systems; x1max – the maximu m nodes number in a distributed computer system; x2max – the maximu m number of data link channels; x3max – the maximu m nu mber of the active security mechanisms; x4max – the maximu m nu mber o f the security monitoring tools in the distributed computer systems.
I. INT RODUCT ION There are a significant number of vulnerabilities in the Distributed Co mputer Systems (DCS), wh ich can be used by the intrusions agents for the attacks launching to get the unauthorized access to the information. In turn, there are known a nu mber of methods and mechanisms for the intrusion detection and the appropriate host and networkbased systems for the monitoring and intrusion prevention [1]. ІSO15408 [2], [3] defines the requirements to the security mechanism and for the risk analysis at various stages of the security mechanisms design, development, imp lementation and maintenance. Also, this standard defines the function of the DCS owners and the informat ion that should be implemented for the protection against possible threats and intrusions. The security risk analysis for the DCS is one of the most important trends in the field of modern effect ive mechanisms for the information protection [4], [5]. The security of the DCS resources is based on their protection from the threats that are classified in view of the possible level of safety violation [6][9]. We consider all kinds of threats, and the most dangerous are those that are performed by the subjects. Fig. 1 shows the main components of the DCS security and their interconnection [1][3].
I.J. Intelligent Systems and Applications, 2016, 12, 5764
58
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
Intrusions Agents
Threats
Vulnerabilities
Risks
Security mechanisms
Resources
Resources owners Fig.1. T he components of the DCS security and their interconnection .
The effective protection of the modern co mputer systems requires support the safety of the operations with their resources, so there is a necessity for the timely identification of the potential threats to their security. It is necessary to analyze the fu ll range of the possible threats. [1], [10][12]. The typical violations of the DCS security are [1][3]: • disclosure of the resources confidentiality that leads to the certain damage; • vio lation of the resources integrity as a result of unauthorized modification; • b locking of the access to the resources, in particular due to incorrect ban for the legal subjects. The resources of the co mputer systems have a certain value, in particular, for their owners and for the intrusion agents as well. So, the security threats have potentially adverse effects on DCS resources, resulting in, inter alia, reducing the resources value [1]. An analysis of existing threats allows us to define the risks to the security of the in formation systems, in addition, this analysis allo ws us to determine the countermeasures to neutralize potential threats and reduce security risks to an acceptable level. Standard ІSO15408 defines the requirements to the risk analysis procedure at the various stages of design, development, imp lementation and maintenance of the security mechanis ms, as well as the actions of the owners of information and informat ion systems admin istrators, to protect them fro m possible threats and intrusions [2], [3]. Countermeasures can reduce the number of potential vulnerabilities in the information system with the protection mechanisms applying in accordance with established security policy. At the same time in case when the countermeasures are already implemented there Copyright © 2016 MECS
may remain the residual vulnerabilities, wh ich can be used by intrusions agents. These vulnerabilities form a residual risk, wh ich is mit igated by the use of additional security mechanisms. The owners need to make sure that the countermeasures employed by them prov ide an adequate reaction to the potential threats before to grant the access to their resources. In general, resource owners may not always objectively evaluate the effectiveness of the security mechanisms that are used and in this case there is required an independent assessment. The result of this assessment is the trust level to the security mechanisms in terms of reducing the risk of resources security. The trust level is characteristics of protection, which confirms the correctness and efficiency of their use, as this rating is used by the resource owner in the decision making process how to set an acceptable level o f security risk to its resources [13]. The results of evaluation of security mechanis ms parameters must be verified in terms of their correctness, which will allow use the received assessment data to substantiate an acceptable level of security risk for a certain information system [14].
II. M ET HODS FOR T HE EFFECT S PREDICT ING OF T HE DANGEROUS A CT IONS OF T HE INT RUSION A GENT S The methods for the evaluation and prediction of consequences of the dangerous actions of intrusion agents by the time are divided into two groups [15][18]: • methods, that are based on a priori estimates obtained on the basis of theoretical models and analogies; • methods, that are based on a posteriori assessments the effects assessments for the already realized intrusion into computer systems. I.J. Intelligent Systems and Applications, 2016, 12, 5764
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
According to the initial info rmation, the methods for predicting of the dangerous actions consequences are divided into [17][19]: • experimental: based on the processing of the parameters of the realized intrusion; • co mbined: co mputational and experimental, based on the statistical data processing with the special mathematical models; • mathemat ical: based on the mathematical models exclusively. The estimated models, that are used to get a priori estimates, perform the tests on the really realized attacks and intrusions into computer systems. A priori estimates of the effects of the hazardous actions differ by the time and goal [17][19]: • preliminary assessment of the various scenarios of the dangerous actions initiating, performed by the preventive mechanis ms planning to protect the computer systems: it provides a basic security level, security monitoring, operative security control; • operational parameter estimates of the realized intrusions performed to get an adequate reaction to security incidents. The evaluation and prediction of the user actions involve the collection and processing of the data on the hazardous activities of the intrusion agents, the definition the perimeter of the co mputer systems components, that are under attack potentially, determination of the impact of the negative factors on the computer system functioning [20], [21]. This assessment allows select the most effective ways for the co mputer system protection, minimizes losses due to the attacks realization. [20], [21]. The risk assessment of the dangerous actions realization, i.e., the analysis of the dangerous actions repeatability and the alleged damage to a certain computer system is performed periodically, in particular, when there is performing an audit of the co mputer security for the risks control [22]. During the assessment of the dangerous actions realization, all the main risk factors are uncertain and there are used their estimates. The forecast of the consequences of the dangerous actions of intrusions agents is a preliminary forecast for the certain computer system and for the certain type of the dangerous actions. The preliminary assessment of the impact of the hazardous activities is a particular problem of the risk assessment in a case when the initiating event has already occurred, i.e. the dangerous action was realized [21], [22]. The fo recast is performed on the uncertain factors parameters, and the results of the prediction are used in the planning of the preventive mechanisms to protect the computer systems [22], [23]. The mathematical models for the consequences prediction of the dangerous actions of intrusion agents are based on a probabilistic approach assuming that the intrusion event is occurred already [24]. This takes into account both the probabilistic nature of the negative Copyright © 2016 MECS
59
factors impact on the co mputer system, and the vulnerabilities of the co mputer system to these influences [24]. The levels of impact factors on the computer system security are the rando m variables and are specified by the relevant distribution. The vulnerab ilit ies of the co mputer system co mponents and resources are also described as random factors, due to the specifics of the development and support of the hardware and software. The random factors affecting the dangerous actions consequences are related to the danger degree, to the spacetime threats and to the vulnerabilities of the computer system [24], [25]. The emergency evaluation of the situation in the case of the dangerous actions is performed by the time parameters of intrusions attempts, by the type of the computer system component that is under the attacks, and by the danger degree of the intrusion actions, that is received fro m the security mon itoring and control subsystems [24], [25]. The emergency assessment is a special case of the abovedescribed problem in case if the hazards and threats have already been realized. The results of this assessment are used in decisions making for the security control of the computer systems, to perform an effective reaction to security incidents [18]. The assessment of the real situation that has arisen as a result of dangerous actions is performed by the data, that are obtained fro m safety mon itoring subsystem, and this assessment allow eliminate the uncertainty by the only one remain ing uncertain riskfactor  the losses [18], [19]. The results of this assessment are used to clarify the earlier decisions on the implementation of the protective mechanis ms and to minimize the consequences of the intrusions. The effect ive risk control imp lies the risks factors prediction, i.e. there are estimated the possible risks indicators for the some future time interval, for examp le, based on the methods of the time series extrapolation, autoregression, and the others [25], [26]. In order to perform the risk pred iction, the special mode ls are used, which describe the prospective state of DCS based on an analysis of its behavior during operations in the previous time intervals [26]. Let us present the specifics of the risk p rediction task in apply to the security risk forecast for the dis tributed computer systems, which is solved to control of the DCS security. The following methods can be used in the risk prediction process [1], [2], [17][19]: • the prediction of the possible safety violation, i.e. the realization of the corresponding triggering dangerous events; • the forecasting of the effects of hazardous events, i.e., the safety violation. For a timely forecasting of hazardous events during the operation of the computer system or to identify them at the initial stage there is required a special system fo r the subjects actions monitoring in the computer systems. Based on the information, that is received on the monitored data, the security administrator makes the operative decisions on the security mechanisms activation I.J. Intelligent Systems and Applications, 2016, 12, 5764
60
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
to prevent and/or to reduce the loss due to a security violations (attacks) on computer systems [17][19]. The methods for the forecast of the dangerous events by the projected parameters, in turn, are subdivided into methods of forecasting their location, level, the o nset time and the repetition rate [27][30]. This problem can be solved in different ways depending on the specifics of the computer systems. The most effective approach is to proactively predict the dangerous events (attacks) and to apply the adequate measures in the stage of their preparation [17][19]. In the case when the time is limited for an adequate reaction to a hazardous event, it is expedient to activate the preventive security mechanis m, based on the statistical evaluation of the recurrence frequency of dangerous events [1]. In fact, the task of the security risks analysis is associated with the assessment of the risk of exceeding of a predetermined maximu m t ime interval for the reaction Network security mechanisms
to possible intrusions in the DCS Job Manager (JM) functions [34]. This risk depends on two factors: the probability and the consequences of the exceeding of the time for the situation analysis and decision making in the DCS JM. The p robability of the time exceeding is functionally linked to the values of the upper and lower bounds of the allowable t ime period for the situation analysis and decisionmaking [32]. In general, the follo wing factors have impact on the security risks: the nu mber of nodes and link channels of DCS, the intensity of the security threats to DCS resources realization, the intensity of security mechanis ms activation, i.e., the rate of reaction to the hazardous events [1], [17][19], [31], [32]. In the case, when the rate of the security threats realizat ion is exceeded the reaction rate, the attack is launched on the DCS resources [33]. Local node security mechanisms
Network monitoring mechanisms
Local node monitoring mechanisms
Local node Queue of network tasks
Queue of local tasks
Block of situation analysis Blockanalyzer on the high level
Block for decision making on the control
Blockanalyzer on the low level
FS
Local job manager
Network job manager Block of access control
RAM
Block of access control
Block for decision making on the control
Block of situation analysis Blockanalyzer on the high level Blockanalyzer on the low level
Network environment for data transfer
Node1
Node2
Node3
Nodei
Nodej
Noden
Fig.2. T he generalized structural scheme of Job Manager in the context of riskbased approach to security analysis (RAM – Random Access Memory, FS – File system)
Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 5764
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
III. THE DCS JOB M ANAGER IN T HE CONT EXT OF RISKBASED A PPROACH FOR SECURIT Y A NALYSIS The analysis of the existing security threats allow determine the risks of safety violations for the co mputer systems resources and determine the countermeasures to neutralize the potential threats, by reducing the security risks to an acceptable level [31]. Taken countermeasures allo w reduce the number of potential vulnerab ilit ies in the computer system, and for this purpose, in accordance with the current safety policy in DCS, are imp lemented the protection mechanisms. However, even if the countermeasures are already implemented there may remain socalled residual vulnerabilit ies that form the residual security risk, wh ich can be reduced by the additional protective mechanisms [31]. One of the most important components of modern distributed computer systems is the Job Manager, which, inter alia, performs the distribution of solving tasks between the DCS resources, including the tasks related to the DCS security. The malfunctions of JM, particularly the time delays during the analysis of critical situations and decisions making, especially in the case when the DCS process the confidential in formation can disrupt or even a completely stop the DCS functioning [31], [32]. Thus, the parameters analysis and the reducing of security risk for Job Manager operations is a very important task. Fig. 2 shows a generalized structural scheme of Job Manager [35][37] in the context of a riskbased approach.
IV. RISKBASED SECURIT Y A NALYSIS OF T HE DIST RIBUT ED AND SCALABLE COMPUT ER SYST EMS Let us introduce the next parameters: x1  the nu mber of nodes in a distributed computer system; x2  the number of data link channels; x3  the nu mber of the active security mechanisms; x4  the number o f the security monitoring tools in the DCS. Next , let α  the average intensity of the security threats for the DCS resources x1 and x2 ; β – the average intensity of the security mechanisms fro m the group x3 and x4 activation, and the parameter R0 – the initial security risk. To assess the security risks in the DCS, we introduce a function:
R
R(t ) 1 e
0 x3 x4 x1 x2 ( )t x3 max x4 max x1 max x2 max
(1)
Function R(t) is a risk of the DCS security threats realization, in depending on the time t. There are some various ways for the risks changing, which depend on the relationship between the expressions:
x3 x 4 x1 x2 and . x1 max x2 max x3 max x4 max
Copyright © 2016 MECS
(2)
If
61
x3 x4 x1 x2 , then where the x1max x2 max x3 max x4 max
security risks are growing and threats can be transformed into a real attack. Conversely: if
x3 x4 x1 x2 , then there is a x1max x2 max x3 max x4 max
reduction of the security risks and the possibility of the attacks realization is low. According to the statistical data, in the modern distributed computer systems, an average 25 intrusion attempts are launched per workstation in the daytime period [18]. Then in the DCS, consisting of 1000 nodes workstations, there are performed about 25,000 intrusion attempts per daytime period, and thus the expected value of the intensity of the security threat realization is αi =0.3, i.e., on average, the 1 intrusion attempt is launched every 3.33 seconds. The evaluations of the security risks R in DCS allow performing the simulat ion for different configurations of distributed computer systems considering the resources scaling. Let us consider the configuration of a distributed DCS, which after the scaling has the following parameters: the maximu m nodes number x1 max=1000; the maximu m nu mber of data lin ks channels x2max = 5500; the maximu m nu mber of active security mechanisms x3 max = 100; the maximu m nu mber of active security monitoring tools x4max = 50. Let us consider the various configurations of the DCS: 1. The DCS is scaled to the largest possible number of nodes and data links channels (x1max = 1000, x2ma x = 5500), and there are the maximu m possible number of active security mechanism and active security monitoring tools (x3max = 100, x4max=50). 2. The DCS is used x3 = 50 active security mechanis ms; x4 = 20 active security mon itoring tools for the maximu m possible number o f nodes and data links channels (x1max = 1000, x2max = 5500); 3. The DCS is used the maximu m possible number of active security mechanis ms and active security monitoring tools (x3 max = 100, x4 max = 50), and the DCS have x1 = 700 nodes and x2 = 3700 data link channels; 4. The DCS is used x3 = 50 active security mechanis ms; x4 = 20 active security monitoring tools, and the DCS have x1 = 700 nodes and x2 = 3700 data link channels. Let us fix an average intensity of activation of the security mechanis ms fro m the group x3 and x4 at the level β = 0.4, the average intensity of realizat ion of the security threats to DCS resources x1 and x2 at the level α = 0.3. Next, we draw a set of the curves that are reflecting the functional dependence between the risk value R and the time t for the four above mentioned DCS configurations with the init ial risk R0 = 0.2. The results are shown in Table 1.
I.J. Intelligent Systems and Applications, 2016, 12, 5764
62
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
T able 1. T he values of the security risks R in DCS t,Ń X1max = 1000, x2max = 5500, x3max = 100, x4max = 50 X1max = 1000, x2max = 5500, x3 = 50, x4 = 20 X1 = 700, x2 = 3700, x3max = 100, x4max = 50 X1 = 700, x2 = 3700, x3 = 50, x4 = 20
0
1
2
3
4
5
6
7
8
9
10
0.2
0.190
0.180
0.170
0.161
0.151
0.142
0.133
0.124
0.116
0.108
0.2
0.222
0.243
0.264
0.283
0.300
0.316
0.329
0.341
0.351
0.360
0.2
0.174
0.149
0.126
0.105
0.086
0.07
0.056
0.045
0.036
0.028
0.2
0.206
0.212
0.218
0.224
0.230
0.236
0.242
0.248
0.254
0.259
In accordance with the data fro m the table 1, the Fig. 3 shows the dynamics of security risk R in time t. Fig. 3 shows that in the DCS configurations 2 and 4 there are the increasing of the security risk, that is caused by the fact that the number of active security and monitoring mechanis ms is insufficient for a given amount of the DCS resources and for a given intensities of intrusion realization and the security mechanisms activation. On the other hand, the for the DCS configurations 1 and 3,
the number of active security and monitoring mechanisms is quite sufficient for preventive and protective measures, accordingly, the security risk is reduced. In the DCS configuration 3 the security risk is reducing in a sufficiently fast way and the risk level 0.15 is achieved after 3 seconds after the activation of the security mechanis ms. Fo r the DCS configuration 1, the same reduction of the security risk level takes about 6 seconds.
R
0.4
DCS 2
0.35 0.3 DCS 4
0.25 0.2
0.15 DCS 1
0.1 0.05
DCS 3
0 1
2
3
4
5
6
7
8
9
10
11
t
Fig.3. T he dynamics of security risk R in time t.
This case is explained by the fact, that there is the excessive number of the active security and monitoring tools for DCS configuration 3. However, such redundancy allows in a more fast way provide the secure Copyright ÂŠ 2016 MECS
state of the computer system, thereby to min imize the time interval when the DCS is in a potentially dangerous state, that reduces the probability of the successful attacks from the intrusion agents. I.J. Intelligent Systems and Applications, 2016, 12, 5764
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
V. CONCLUSION Thus, the proposed analytical assessment for security risk level in the DCS allo ws forecast the values of the critical time fo r the situation analysis and decision making for the current configuration o f a distributed computing system. Th is forecast takes into account the number of used nodes and data links channels, the number of act ive security and monitoring mechanisms at the current period, as well as on the intensity of the events related to the security threats and the activation intensity of the intrusion prevention mechanisms. Also, the proposed comprehensive analytical assessments of the risk allo w analy zing the dynamics of intrusion processes, the dynamics of the security level recovery and the corresponding dynamics of the risks level in the DCS. The analysis of the probability of a crit ical t ime exceeding for the decisionmaking, i.e. the reaction t ime of the security mechanisms to the threats in case of dynamic change of DCS security parameters, requires to perform a formal description of the processes in the DCS job manager, in part icular, with dedicated networking mechanis ms and models that will allow dynamically evaluate the interval boundaries of critical response time for the actions of the intrusions agents. REFERENCES E. M aiwald, “Fundamentals of network security,” Technology Education, 2004. [2] ISO / IEC 1540812002 Information technology. M ethods and security features. Criteria for Information Technology Security Evaluation. [3] ISO / IEC 154082009 Information technology  Security techniques  Evaluation criteria for IT security. http://ebookbrowse.com/i/iso154081pdf [4] Risk M anagement Guide for Information Technology Systems: SP 80030. – Recommendations of the National Institute of Standards and Technology, 2002. [5] ISO/IEC 27001:2005 Information technology  Security techniques  Information security management systems – Requirements, 2005. [6] E. Alsous, A. Alsous, and A Botnet, “Detection System Using M ultiple Classifiers Strategy ,” International Review on Computers and Software, Vol. 7. n. 5, 2012, pp. 20222028. [7] K. Suresh Kumar, and T. Sasikala, “A Technique for Web Security Using M utual Authentication and ClickingCropping Based Image Captcha Technology ,” International Review on Computers and Software, Vol. 9, n. 1, 2014, pp. 110118. [8] R. Pal, and P. Hui, “Cyberinsurance for cybersecurity: A topological take on modulating insurance premiums. Performance Evaluation Review,” 2012. [9] K. Rama Abirami, M . G. Sumithra, and J. Rajasekaran, “An Efficient Secure Enhanced Routing Protocol for DDoS Attacks in MANET,” International Review on Computers and Software, Vol. 9, n. 1, 2014, pp. 119 – 127. [10] S. I. Sabasti Prabu, and V. J. Senthil Kumar, “Entropy Based Approach to Prevent the DDoS Attacks for Secured Web Services,” International Review on Computers and Software, Vol. 8. n. 4, 2013, pp. 888891. [11] “Practical Threat Analysis for Information Security Experts,” TA Technologies, 2010. [1]
Copyright © 2016 MECS
63
http://www.ptatechnologies.com/default.htm [12] Y. S en ha j i, H . M edr om i, an d S. T a ll a l, “Network Security: Android Intrusion Attack on an Arduino Network IDS,” International Review on Computers and Software, Vol 10, n. 9, 2015, pp. 950958. [13] V.M ukhin. Adaptive Approach to Safety Control and Security System M odification in Computer Systems and Networks, Proceedings of the 5th IEEE Workshop on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS’2009), Rende (Cosenza), Italy, 21  23 September 2009. – pp. 212  217. [14] V. M ukhin, A Bidkov, Vu Duc Thinh. The Forming of Trust Level to the Nodes in the Distributed Computer Systems, Proc. of XIth International Conference “Modern Problems of Radio Engineering, Telecommunications and Computer Science TCSET’2012”, Lvov – Slavsko, 21 24 February 2012. – p. 362. [15] IEC/ISO 31010 Risk management – Risk assessment techniques, 2009. [16] ISO 31000:2009 Risk management  Principles and guidelines, 2009. [17] P. Hopkin, “Fundamentals of Risk M anagement: Understanding, Evaluating and Implementing – The Institute of Risk M anagement,” 2010. [18] R. Sohizadeh, M . Hassanzadeh, H. Raddum, and K. Hole, “Quantitative risk assessment,” 2011. [19] “Risk M anagement Fundamentals Homeland Security Risk M anagement Doctrine,” 2011. https://www.dhs.gov/xlibrary/assets/rmariskmanagementfundamentals.pdf [20] “Security risk analysis and management”. https://www.nr.no/~abie/RA_by_Jenkins.pdf [21] “Risk management, concept and methods,” CLUSIF, White paper. https://www.clusif.asso.fr/fr/production/ouvrages/pdf/CL USIFriskmanagement.pdf [22] “M anaging Information Security Risk Organization, M ission, and Information System View,” NIST Special Publication 80039, M arch 2011. http://csrc.nist.gov/publications/nistpubs/80039/SP80039final.pdf [23] “Basics of Risk Analysis and Risk M anagement”. http://www.hhs.gov/sites/default/files/ocr/privacy/hipaa/a dministrative/securityrule/riskassessment.pdf [24] “NHS Information Risk M anagement Digital Information Policy NHS Connecting for Health,” 2009. http://systems.hscic.gov.uk/infogov/security/risk/inforisk mgtgpg.pdf [25] M . Ketel, “It security risk management,” Proceedings of the 4 6th Annual Southeast Regional Conference, 2008. [26] L. P. Rees, J. K. Deane, T. R. Rakes, and W. H. Baker, “Decision support for cybersecurity risk planning,’ Decision Support Systems, vol. 51. no. 3. 2011, pp. 493505. [27] P. Saripalh, and B. Walters, “Quire: A quantitative impact and risk assessment framework for cloud security ,” IEEE 3rd International Conference on Cloud Computing, 2010. [28] M ohamed Hamdi, and Noureddine Boudriga, “Computer and network security risk management: theory, challenges, and countermeasures,” International Journal of Communication Systems, Volume 18, Issue 8, 2005, pp. 763–793. DOI: 10.1002/dac.729 [29] Yang Liu, Zhikui Chen, and Xiaoning Lv, “Risk computing based on capacity of riskabsorbing in virtual community environment,” International Journal of Communication Systems, 2014.
I.J. Intelligent Systems and Applications, 2016, 12, 5764
64
Analytical Assessment of Security Level of Distributed and Scalable Computer Systems
[30] Shi, C. Beard, and K. M itchell, “Analytical M odels for Understanding M isbehavior and M AC Friendliness in CSM A Networks,” Performance Evaluation, Vol. 66 (9– 10), 2009, pp. 469. DOI:10.1016/j.peva.2009.02.002. [31] N. M ohammadi, and M . Zangeneh, “Customer Credit Risk Assessment using Artificial Neural Networks,” I.J. Information Technology and Computer Science, Vol.8, N3, 2016, pp. 5866. DOI: 10.5815/ijitcs.2016.03.07 [32] P. R. Vamsi, and K. Kan, “Self Adaptive Trust M odel for Secure Geographic Routing in Wireless Sensor Networks,” International Journal of Intelligent Systems and Applications, Vol. 7, N3, 2015, pp. 2128. DOI: 10.5815/ijisa.2015.03.03 [33] A. Koul, and M . Sharma, “Cumulative Techniques for Overcoming Security Threats in M anets,” International Journal of Computer Network and Information Security, Vol. 7, N. 5, 2015, pp.6173. DOI: 10.5815/ijcnis.2015.05.08 [34] S. G. Ponnambalam, P. Aravindan, and P Sreenivasa Rao, “Comparative equation of genet IC algorithms for jobshop scheduling,” Production Planning Control, vol. 12, 2001, pp. 560574. [35] S. Dimassi, A. B. Abdelali, A. M rabet, M . N. Krifa, and A. Mtibaa, “A M odeling Tool for Dynamically Reconfigurable Systems,” International Review on Computers and Software, Vol. 9, n. 4, 2014, pp. 600608. [36] H. Shan, W. Smith, L. Oliker, and R. Biswas, “Job scheduling in a heterogeneous GRID environment”. http://www.osti.gov/bridge/servlets/purl/860301UfJBKm/860301.pdf [37] I. Bow ma n , “Conceptual Architecture of the Linux Kernel”. www.grad.math.uwaterloo.ca/~itbowman/CS746G/a1/
Authors’ Profiles Zhengbing Hu: Associated Professor of School of Educational Information Technology, Huazhong Normal University , PhD. M . Sc (2002), PhD (2006) from the National Technical University of Ukraine Kiev Polytechnic Institute”. Postdoctor (2008), Huazhong University of Science and Technology, China. Honorary Associate Researcher (2012),Hong Kong University, Hong Kong, M ajor interest: computer science and technology applications, artificial intelligence, network security , communications, data processing, cloud computing, education technology.
Vadym Mukhin: Professor of computer systems department of National Technical University of Ukraine “Kiev Polytechnic Institute”, Doct. of Sc. Born on November 1, 1971. M . Sc. (1994), PhD (1997), Doct. of Sc. (2015) from the National Technical University of Ukraine “Kiev Polytechnic Institute”; Assoc. Prof. (2000), Professor (2015) of computer systems department. M ajor interest: the security of distributed computer systems and risk analysis; design of the information security systems; mechanisms for the adaptive security control in distributed computing systems; the security policy development for the
Copyright © 2016 MECS
computer systems and networks.
Oleg Barabash: Professor of networking and Internet technologies department, Taras Shevchenko National University of Kiev, Ukraine, Kiev, Doct. of Sc. Born on July 28, 1964. M. Sc. (1986), PhD (1992), Doct. of Sc. (2006) from the National Academy of Defence of Ukraine; Assoc. Prof. (1996), Professor (2007) of computer systems department. M ajor interest: the functional stability of information systems and diagnostic systems for digital objects; security of distributed computer systems; design of the information security systems computer systems; design of the information security systems.
Yaroslav Kornaga: Assoc. professor of computer systems department of National Technical University of Ukraine “Kiev Polytechnic Institute”, PhD. Born on January 1, 1982. M . Sc. (2005), PhD (2015), from State University of Telecommunications; Assoc. Prof. (2015) of techical cybernetics department. M ajor interest: the security of distributed database and risk analysis; design of the distributed database; mechanisms for the adaptive security control in distributed database; the security policy development for distributed database.
Oksana Herasymenko Assist. professor of networking and Internet technologies department, Taras Shevchenko National University of Kiev. Born at 1983 in Borzna town, Chernihiv region, Ukraine. M . Sc. in computer system and networks (2006) from Chernihiv State Technological University, Ukraine. M ajor interests: resource control in distributed computing system and data mining. Now she is preparing her Ph.D. thesis in information technologies.
Yaroslav Lavrenko: Assoc. professor of dynamics and strength of machines and strength of materials department of National Technical University of Ukraine “Kiev Polytechnic Institute”, PhD. Born on M arch 15, 1983. M . Sc. (2006), PhD (2014), from National Technical University of Ukraine “Kiev Polytechnic Institute”; Assoc. Prof. (2015) of dynamics and strength of machines and strength of materials department. M ajor interest: Dynamics and durability of high rotation system, vibrations.
How to cite this paper: Zhengbing Hu, Vadym M ukhin, Yaroslav Kornaga, Yaroslav Lavrenko, Oleg Barabash, Oksana Herasymenko, "Analytical Assessment of Security Level of Distributed and Scalable Computer Systems", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.5764, 2016. DOI: 10.5815/ijisa.2016.12.07
I.J. Intelligent Systems and Applications, 2016, 12, 5764
I.J. Intelligent Systems and Applications, 2016, 12, 6572 Published Online December 2016 in MECS (http://www.mecspress.org/) DOI: 10.5815/ijisa.2016.12.08
Empirical Study of Impact of Various Concept Drifts in Data Stream Mining Methods Veena Mittal FET (CSE), MRIU, Faridabad Email: veena.mittal06@gmail.co m
Indu Kashyap FET (CSE), MRIU, Faridabad Email: indu.fet@mriu.edu.in.
Abstractâ€”In the real wo rld, most of the applications are inherently dynamic in nature i.e. their underlying data distribution changes with time. As a result, the concept drifts occur very frequently in the data stream. Concept drifts in data stream increase the challenges in learning as well, it also significantly decreases the accuracy of the classifier. However, recently many algorith ms have been proposed that exclusively designed for data stream mining while considering drift ing concept in the data stream.This paper presents an empirical evaluation of these algorith ms on datasets having four possible types of concept drifts namely; sudden, gradual, incremental, and recurring drifts.
often. Concept drifts occurred when the concepts represented by the continuously collected data changes with time after having a minimum stability period [3].
Index Termsâ€”Concept drift, on line learn ing, data stream mining, drift detection, ensembles.
I. INT RODUCT ION Learn ing classifiers fro m train ing datasets is one of the most important steps in data mining and machine learn ing. Until now, many algorith ms that are used for learn ing classifiers are based on the static environment of underlying data distribution, which remains static and does not change with time. In this scenario, the complete data can be stored in memory electronically due to which, it is possible to process the data several times. However, in many applicat ions, such as weather forecasting, traffic management, sensor network, etc.[3] the underlying data distribution changes very frequently, and it is usually beyond the capacity of traditional s tatic learning algorith ms to work accurately in such dynamic environment too. In the dynamic environ ment, the generated data exh ibits the characteristics of data streams. The data streams are categorized by its frequent generation rate and big data volu mes , which requires a fast response in a manner to make decisions in real t ime. In contrast to algorithms designed for learn ing in a static environment, the learning algorith ms of data streams with changing concepts should require fu lfilling some new constraints such as one pass testing, memo ry limitat ions and time constraints [12],[ 56]. Furthermore, in data streams the change in targeted concept with time called concept drift [4], [6] is quite Copyright ÂŠ 2016 MECS
Fig.1. Various types of drifts in data streams
Frequent occurrences of concept drifts in the data streams decrease the performance of the classifiers significantly. The concept drifts is broadly categorized into follo wing categories (i) sudden (ii) gradual (iii) incremental and (iv) recurring, as shown in Figure 1. In the presence of concept drifts, a good classification algorith m should be able to adapt itself to cater the changes in underlying data distribution in a manner to achieve the consistent accuracy of the classifier during classification of unseen instances that are arriv ing continuously with time. Many incremental algorith ms that learn incrementally over the data needed to update for every new unseen instance during learning. However, to deal with concept drifts, it is necessary for the incremental algorith m that it should ensure the forgetting of old concepts and quick I.J. Intelligent Systems and Applications, 2016, 12, 6572
66
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
adaptation to new concepts [3]. Recently, many methods have been proposed in the related literature that exclusively designed for data stream mining while considering drift ing concept in the data stream. They are categorized as online learn ing algorith ms and mainly include sliding windowbased methods, ensembles, and drift detection methods. These methods fulfill the one pass requirement of learn ing in data stream without storing the data electronically. The online approaches can be broadly discussed in two categories: (i) Online learning approaches that use an explicit mechanis m to deal with concept drifts [7], [9],[10] and (ii) On line learning approaches that do not use any exp licit mechanism to deal with concept drifts [1115][32]. Most popularly, the former online learn ing approaches include Early Drift Detection Method (EDDM) [7] and Drift Detection Method (DDM ) [10]. The later approaches basically emp loy a set of learners also called ensembles are used in which each learner is assigned some weight, depending on the accuracy of learners. Ensemb les are popularly used to increase the accuracy in static data problem. However, they need certain modifications to justify their applicability in datasets with changing environment. In general, consistent updating of ensemble structure and weights of learners is required for adopting the change in the environment. Inthis paper, we present an empirical evaluation of some popular algorith ms , e.g., standard NaĂŻ ve Bayesian (NB), Drift Detection Method (DDM ) [10], Weighted Majority Algorithm (WMA) [8], [16], Accuracy Updated Ensemb le (AUE), Hoeffding Option Tree (HOT) on artificialdata stream mining with drifting concepts using datasets having four possible types of concept drifts namely: sudden, gradual, incremental, and recurring drifts. The rest of the paper organized as follows.Section II illustrates the related terminologies and concepts related to data stream min ing withconcepts drift ing, this section also presents related work in the area of data,Section III describes experimental setup and datasets that we have used for emp irical evaluation of algorith ms. Finally, Section IV discusses theresults and section V provide the conclusion of overall empirical findings.
II. RELAT ED W ORK A. Concepts and terminologies related to data stream mining Let represents a dimensional instance of data stream training dataset at time t, where t represents the { } time such that and where represents the class of data instance . Therefore, wh ile considering the problem of data stream mining as the supervised incremental learning process, the task is to predict the class of new t rain ing instance , if the predicted is the same as actual class of the instance, it is assumed that classifier is working well and if the prediction is wrong then updating of learning algorith m is mandatory. Such approach of training the Copyright ÂŠ 2016 MECS
classifiers for data stream min ing isalso called Prequential method of learning. Another approach to learning is called batch learning. There are various alternates available for perfo rming batch learn ing. One of the most popular approaches for batch learn ing is to divide the complete training dataset into equal sized data batches, such that , where, represents the ith batch.
Fig.2. Various approaches for dealing with concept drifts in data streams
Concept drifts are very often in data streams, concept drifts generally occur due to change in underly ing distribution, and there are many approaches described in related literature to detect the drifts [7]. As already described, the drifts can be main ly categorized into sudden, gradual, recurring and incremental. Sudden drifts occur when underlying data distribution changes suddenly, whereas in the gradual drifts , the distribution changes radically. In incremental drifts, the change in drift ing concepts are very small, but persist for a long time. As a result the resultant the class change occurs completely after a long time. In the recurring type of concept drifts, the concepts keep repeating time to time. Based on the basic techniques, which have been used in the various learning algorith ms, we can alternately categorize the methods for dealing with concept drifts in data stream min ing in three different categories i.e., (i) algorith ms based window technique (ii) algorith ms based on drift detection methods, and (iii) algorith ms based on ensembles methods. The sliding windowbased approaches are very common in data stream min ing [17]. In slid ing windowbased methods, it is very important to decide an ideal window size. A s mall window size ensures the fast response to drift; however, it is very often to have false I.J. Intelligent Systems and Applications, 2016, 12, 6572
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
detection also. On another hand, a large window size suffers fro m delay in detection. For these reasons, dynamically ad justed window methods are well suited in utilizing windowbased methods in data stream min ing [17]. Alternately, the algorith ms that employ the drift detection methods are based on the consistent statistical observation on the change in the class distribution. If any change occurs in the distribution, then base classifier is reconstructed to manage the change [18]. Drift Detection Method (DDM) [10] is one of the most popular approaches among the algorithms that uses drift detection in data stream min ing. In DDM, drift detection is performed by monitoring the predict ion error, wh ich is modeled as bino mial distribution. If the error rate lies beyond the decided value; an alarm is generated as an action, the current classifier is dropped and a new classifier is constructed. DDM performs well for sudden drifts comparative to the gradual and incremental. Another approach is suggested in many kinds of literature that has been follo wed by the many algorith ms of data stream mining is called ensemble method, wh ich is quite apart from window and drift detection methods. Furthermore, the ensemble approaches can further be classified in approaches that incrementally learn fro m each coming instances one by one online and ensembles that learn in batches. Diversity among the base learners is the main issue of concern in ensemble methods; this necessity can be ensured by using online bagging [19] in which base learners are t rained incrementally, and decisions of learners are co mbined using majority weights. Leverag ing bagging [20] adds more randomizat ion to the bagging method. The DDD algorith m [21]analyzes the effect of diversity in ensembles by comb ining four different d iverse ensembles. B. Description of evaluated algorithms This section presents the description of all algorithms, which we have used for performing our analysis on various types of concept drifts. i.
Naïve Bayesian (NB)
NB classifiers are one of the most popular probabilistic classifiers based on Bayes theorem [22], [23], [24].The inherited property of NB makes it a good streaming method which suits well dynamic environment though its success in ensemble technique is in [11]. ii.
Drift Detection Method (DDM)
Drift Detection Method (DDM ) [11], as mentioned above is based on drift detection, when a drift is detected the system rebuild itself to incorporate the change in the concept. In DDM, the occurrence of drift is traced by monitoring the classification error rate. When the classification error rate reached to the threshold level, the system drops the previous concept and reset itself to learn the new concept. The DDM uses the Binomial Distribution to model the error in classification. The standard deviation ,for each point t, is given in equation (1) using probability of Copyright © 2016 MECS
67
misclassification (p t ) √
⁄
(1)
The standard deviation and minimu m error rate achieved are stored in and DDM, then checks for following conditions: (i) (2) (ii) if (3) The model made by the learning method is rebuilt, and a new model is learned in case (ii) using the stored examples since the warning level occurred. iii.
Weighted Majority (WM)
In machine learning, the WM algorith m [8] is one of the best metalearning algorith ms that is used for ensemble construction. In WM algorithm, each member of the ensemble is init ially assigned the weight of 1 and for every mistake in classification by any member, the weight of the corresponding members decreases by the mu ltip licat ive constant factor of , where the value of .
iv.
Hoeffding option tree(HOT)
Decision tree based classifiers are relat ively fast as compared to other model of classifications. The Hoeffding Tree (HT)[26] is an incremental classifier used for very fast and massive data streams. The HT uses only small subset of training dataset to find the best split. The number of examples required for this is decided b y Hoeffding Bound. Very Fast Decision Tree (VDFT) is an upgraded version of HT that is having refinements on issues like ties, memo ry, co mputation on split function, poor attributes and initialization. Hoeffding Option Tree (HOT) [25] is another variation of decision tree. v.
Accuracy updated ensemble (AUE)
Accuracy Updated Ensemble (AUE) [31], is an extension of AWE. AUE uses online component classifiers, which is updated according to the present distribution. In AUE, addit ional modification in weight function has done to solve the problem with AWE. AUE is more accurate than AWE but required more t ime and memory
III. EXPERIMENT AL SET UP AND EVALUAT ION All experiments are performed using Massive Online Analysis (MOA) [30] framework, where each algorith m is imp lemented using Java language. The co mp lete experiments were conducted on five different art ificial datasets. These datasets are collected from UCI repository, and they are considered as the benchmark for analyzing the data streams. A brief description of the I.J. Intelligent Systems and Applications, 2016, 12, 6572
68
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
collected datasets is given below in Table 1. Description of collected datasets We collected five art ificial datasets from UCI repository [29], the collection of the datasets were made in a manner such that each dataset must have any one of the concept drift namely incremental, gradual, sudden, recurring and mixed. The select ion is made in this manner to have a p roper analysis of various data stream mining algorithms. Hyperplane: Hyperplane is one of most popular dataset generator used for analysis of many algorith ms of data mining [27]. The hyperplane generator is generally used for generation of incremental concept drift. For our experiment, we have set the hyperplane generator to generate 1 million instances of two classes with 10 attributes and one drift only. Please refer to Table 1. Radial Basis Function (RBF):The RBF is the very popular function that has been used in many mach ine algorith ms. Basically, the RBF produces the real values based on a distance called centroid, wh ich is the distance fro m the origin. The RBF produces drifting centroid based on the user input. For our experiments, we have generated 1 million datasets of using RBF generators to produce gradual drifts. Ou r dataset of RBF is having 20 attributes, 4 drifts and 4 classes as shown in Table 1. Streaming Ensemble Algorithm (SEA): For generating sudden drift, we used SEA. A total of 1 million instances were generated with 3 attributes, 4 classes, and 9 drfits. We used MOA to generate this dataset. Please refer to Table 1. Tree: In our experiment we have used the tree dataset, which contains four recurring drifts consistently scattered over 0.1 million instances. Our tree dataset contains 10 attributes, 15 drifts, and 6 classes. Light Emitting Diodes (LED): The LED dataset consists of 24 binary attributes, wh ich defines the digit to displayed over sevensegment display. We used LED function to generate mixed drifts distributed over 1 million instances with 3 drifts and 10 classes. T able 1. Artificial datasets used for experiment
Type s of Drift Incremental
Artificial Dataset Ge nerators
1
No. of Attr ibut es 10
1
20
4
4
1
3
9
4
No. of Instances in millions
No. of Drift s
No. of classe s
1
2
Sudden
Hyperplane Radial Basis Function SEA
Recurring
T ree
0.1
10
15
6
Mixed
LED
1
24
3
10
Gradual
Evaluation All experiments were performed on Massive Online Analysis (MOA) framework, where each algorith m is implemented using JAVA. The experiments carried out Copyright ÂŠ 2016 MECS
on Intel Core i3 (1.8Gh z, 3 MB L3 cache, with 4 GB RAM). We have conducted the experiments for obtaining three wellknown performance measures used for measuring goodness of data stream mining algorithms. These three measures are described below:1.
2.
Prequential Accuracy: The prequential accuracy [28], is the average accuracy of predicting the class of a new instance without learning it, based on the knowledge learned by the previously learned instances. The average prequential accuracy is calculated for a decided window size by taking an average of correctly classified instances in that window. Kappa Statics: Kappa Statics measures the homogeneity among the experts. Ho mogeneity is inversely related to the diversity of the experts i.e. more the homogeneity less the diversity among the experts.
Evaluation Time: Evaluation time is the average time taken by CPU for testing the new instance and training the classifier.
IV. RESULT S A NALYSIS As mentioned in section I, we have conducted the experiments for five different types of algorithm based on three different approaches as shown in figure 2. The graph of figure 37 shows the prequential accuracy of all these five algorith ms on five d ifferent types of datasets that differ in the nature in the drifts respectively. Fro m our experiments, we observed that for incremental type of drifts all algorith ms performed very well with the best performance g iven by DDM and WM with mo re than 90% average prequential accuracy as depicted in the graph of figure 3. The graph of figure 4, depicts the prequential accuracies of all five algorith ms on gradually drift ing dataset generated by radial basis function (RBF). It is very clear fro m the graphs of figure 4, that AUE is performing extraord inarily with average prequential accuracy of more than 93%, while on other hand the DDM and NB perform worst with average prequential accuracy of approx. 73%. Ho wever, the HOT and WM perform co mparatively well with the average prequential accuracy of 90.49% and 89.46 % respectively. The average prequential of all algorith ms for SEA dataset are depicted by the graphs in the figure 5. As we know that the SEA dataset distribution represents the sudden drift, therefo re fro m the graphs it can be easily concluded that the AUE algorith m produces the best prequential accuracy of 89.50 % , which is just follo wed by other algorithms with a very s mall difference. Similarly, in the case of Tree dataset (recurring drift), as shown in figure 6, AUE supersedes all other algorith ms with the average prequential accuracy of 90.6 % just followed by HOT and WM. Ho wever, the performance of the DDM and NB is very low (57.5 %) for the recurring I.J. Intelligent Systems and Applications, 2016, 12, 6572
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
dataset. All algorith ms have shown a comparable performance. The prequential accuracy of various algorith ms in mixed drift dataset is shown in figure 7. Fro m the graphs of the various algorithms in the figure 7, is can be easily determined that all algorith ms are performing mediocre with the average prequential accuracy ranges between 7374 % for all algorith ms. The average prequential accuracy of various algorith ms on various types of datasets is summarized in Table 2. Fro m the table, it can be observed that DDM and NB give the worst performance among the comp lete experiments for recurring drift dataset. The bar graph in figure 8 and figure 9, depicts the average prequential accuracy of algorith ms on different datasets and average prequential accuracy of different algorithms on same dataset respectively.
69
Fig.5. Prequential Accuracy on SEA (Suddenly drifting) dataset
Consistent performers Although, some algorithms are do ing very well on some types of drifts, but on observing the graphs in figure 10, it is very clear that no algorith m is performing consistently very well for all kinds of drifting datasets. However, AUT, HOT and WM are performing mo re consistent as compared to NB and DDM.
Fig.6.Prequential Accuracy on T ree (Recurring) dataset
Fig.3. Prequential Accuracy on Hyperplane (Incremental) dataset
Fig.7. Prequential Accuracy on LED (Mixed Drift) dataset T able 2. Average Prequential Accuracy
Fig.4.Prequential Accuracy on RBF (gradually drifting) dataset
Copyright ÂŠ 2016 MECS
Dataset \Algo. Incremental (Hyperplane) Gradual Drift (RBF) Sudden Drift (SEA) Recurring Drift (Tree) Mixed Drift LED
AUE
DDM
HO T
NB
WM
90.5
93.98
89.53
93.98
93.94
93.14
72.03
90.49
72.02
89.46
89.50
88.22
89.18
88.22
89.30
90.64
57.12
86.99
57.12
87.11
73.88
73.97
73.92
73.97
73.95
I.J. Intelligent Systems and Applications, 2016, 12, 6572
70
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
Fig.8.Average prequential accuracy of algorithms on different datasets. Fig.12. Average CPU time elapsed in seconds for various algorithms on various types of drifting datasets T able 3. CPU T ime elapsed in Seconds Algo/Datasets
Fig.9. Average prequential accuracy of different algorithms on same dataset.
Fig.10. Consistent performance evaluation
Fig.11. CPU time elapsed in seconds for various algorithms on various types of drifting datasets
Copyright ÂŠ 2016 MECS
AUE
Hyperplane 160.78
DDM
6.59
HO T
18.5
NB
3.8
WM
28.8
RBF 138.64
SEA 56.98
Tree 191.02
LED 99.11
8.5
2.39
12.38
5.25
29.48
13.44
36.27
28.14
5.81
1.42
6.41
2.77
31.47
11.19
43.16
19.89
Training Speed of various algorithms The CPU time required by an algorithm is very crucial performance measure fo r determining the goodness of the algorith m in data stream mining of drifting streams. Drift detection delay is an important factor for any algorith m for reacting upon drifts . A large drift detection time ensures more accuracy whereas the small drift detection time reduces the response time but increases the possibilit ies of errors in classification and hence reduces the accuracy of the system. Ideally, a data stream mining algorith m must have small drift detection time with high accuracy. However, the nature of drifts on wh ich algorithm has been trained is also a matter of great consideration. The CPU time elapsed for various algorith ms for various types of drifting datasets in our experiments is depicted by the bar chart of figure 11 and the average time required by an algorithm on the all datasets is given in figure 12. Fro m the bar chart, it can be observed that the AUE algorithm is taking largest CPU time. However, the CPU t ime required by NB and DDM is very less and almost minimu m among all. The A UE algorith m has taken about the largest CPU time of 191.02 seconds for learning in recurring (tree) dataset and minimu m train ing time of 56.98 seconds on sudden (SEA ) drifting datasets. The algorith ms HOT and WM performed quite mediocre as compared to others. The CPU t ime (elapsed in seconds) of various algorith ms on various types of datasets is summarized in Table 3. Kappa statistics As already described the Kappa Statistics measures the homogeneity among the experts. Homogeneity is inversely related to the diversity of the experts i.e. more the homogeneity less the diversity among the experts. The Kappa statics percentage of AUE, DDM, HOT, N B I.J. Intelligent Systems and Applications, 2016, 12, 6572
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
and WM are shown in figure 13 and 14 fo r RBF and Tree datasets respectively.
[3] [4]
[5]
[6]
[7] Fig.13.Kappa statistics of various algorithms on RBF dataset.
[8]
[9]
[10]
[11]
Fig.14. Kappa statistics of various algorithms on T ree dataset.
V. CONCLUSION AND FUT URE SCOPE In our experiments, we examined five different data stream mining algorith ms on five differently drift ing datasets. Fro m our experiments, we have observed that no algorith m is performing uniformly on the categories of the drifting datasets on which we conducted the experiments. It is very crucial for data stream min ing algorith ms to be very fast in detecting the drifts and reset the system appropriately in the response of the change in data distribution. However, in our experiments we observed that it is quite d ifficu lt to maintain a tradeoff between accuracy and CPU t ime. Moreover, it is also observed that no algorith m is uniformly accurate on all kinds of drift ing datasets. Therefore, it is still very challenging task to device an algorith m, wh ich is not only highly and uniformly accurate on every kind of drifts but also fast enough so that it can be compatible with realtime decision making system.
[12]
[13]
[14]
[15]
[16]
[17]
[18] [19]
REFERENCES [1]
[2]
Street, W. Nick, and YongSeog Kim. "A streaming ensemble algorithm (SEA) for largescale classification." Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining. ACM , 2001. Kuncheva, Ludmila I. "Classifier ensembles for detecting
Copyright © 2016 MECS
[20]
[21]
71
concept change in streaming data: Overview and perspectives." 2nd Workshop SUEM A. 2008. Gama, Joao. Knowledge discovery from data streams. CRC Press, 2010. Su, Li, Hongyan Liu, and ZhenHui Song. "A new classification algorithm for data stream." International Journal of M odern Education and Computer Science (IJM ECS) 3.4 (2011): 32. Cao, Yuan, Haibo He, and Hong M an. "SOMKE: Kernel density estimation over data streams by sequences of selforganizing maps." Neural Networks and Learning Systems, IEEE Transactions on 23.8 (2012): 12541268. Sidhu, Parneeta, and M . P. S. Bhatia. "Empirical Support for ConceptDrifting Approaches: Results Based on New Performance M etrics." (2015). Gama, Joao, et al. "Learning with drift detection." Advances in artificial intelligence–SBIA 2004. Springer Berlin Heidelberg, 2004. 286295. Littlestone, Nick, and M anfred K. Warmuth. "The weighted majority algorithm." Information and computation 108.2 (1994): 212261. Nishida, Kyosuke, and Koichiro Yamauchi. "Detecting concept drift using statistical testing." Discovery Science. Springer Berlin Heidelberg, 2007. BaenaGarcıa, M anuel, José del CampoÁvila, Raúl Fidalgo, Albert Bifet, R. Gavalda, and R. M oralesBueno. "Early drift detection method." In Fourth internat ional workshop on knowledge discovery from data streams, vol. 6, pp. 7786. 2006. Kolter, J. Z., & M aloof, M . (2003, November). Dynamic weighted majority: A new ensemble method for tracking concept drift. In Data M ining, 2003. ICDM 2003. Third IEEE International Conference on (pp. 123130). IEEE. J. Z. Kolter and M .A. M aloof‖ Using additive expert ensembles to cope with concept drift‖. In Proceedings of the TwentySecond ACM International Conference on M achine Learning (ICM L’05), Bonn, Germany, pp. 449– 456. J.Z. Kolter and M .A. M aloof, ―Dynamic weighted majority: An ensemble method for drifting concepts‖, JM LR (2007)8: 2755–2790 L. L. M inku and X.Yao, ―DDD: A New Ensemble Approach for Dealing with Concept Drift‖, IEEE Transactions on Knowledge and Data Engineering, VOL. 24, No. 4, 619, 2012. M inku, L. L., White, A. P., & Yao, X. (2010). The impact of diversity on online ensemble learning in the presence of concept drift. Knowledge and Data Engineering, IEEE Transactions on, 22(5), 730742. Blum, Avrim. "Empirical support for winnow and weightedmajorityalgorithms: Results on a calendar scheduling domain." M achine Learning 26.1 (1997): 523. A. Bifet and R. Gavaldà, “Learning from timechanging data with adaptive windowing,” inProc. 7th SIAM Int. Conf. Data M ining, 2007, pp. 443–448. E. S. Page, “Continuous inspection schemes,” Biometrika, vol. 41, nos. 1–2, pp. 100–115, Jun. 1954. N. C. Oza and S. J. Russell, “Experimental comparisons of online and batch versions of bagging and boosting,” inProc. 7th ACM SIGKDD Int. Conf. Knowl. Discovery Data M ining, 2001, pp. 359–364. A. Bifet, G. Holmes, and B. Pfahringer, “Leveraging bagging for evolving data streams,” in Proc. Eur. Conf. M ach. Learn./PKDD, I, 2010, pp. 135–150. L. L. M inku and X. Yao, “DDD: A new ensemble approach for dealing with concept drift,”IEEE Trans. Knowl. Data Eng., vol. 24, no. 4, pp. 619–633, Apr. 2012.
I.J. Intelligent Systems and Applications, 2016, 12, 6572
72
Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods
[22] Friedman, Nir, Dan Geiger, and M oises Goldszmidt. "Bayesian network classifiers." M achine learning 29.23 (1997): 131163. [23] Duda, Richard O., and Peter E. Hart. Pattern classification and scene analysis. Vol. 3. New York: Wiley, 1973. [24] Langley, Pat, Wayne Iba, and Kevin Thompson. "An analysis of Bayesian classifiers." AAAI. Vol. 90. 1992. [25] Bernhard Pfahringer, Geoffrey Holmes, and RichardKirkby. New options for hoeffding trees. In AI, pages 90–99, 2007 [26] P. Domingos and G. Hulten, “M ining highspeed data streams,” inProc. 6th ACM SIGKDD Int. Conf. Knowl. Discovery Data M ining, 2000, pp. 71–80 [27] Fan, Wei, et al. "Active M ining of Data Streams." SDM . 2004. [28] Dawid, A. Philip, and Vladimir G. Vovk. "Prequential probability: Principles and properties." Bernoulli (1999): 125162. [29] UCI M achine Learning Repository archive.ics.uci.edu/ml [30] Bifet, Albert, et al. "M oa: M assive online analysis." The Journal of M achine Learning Research 11 (2010): 16011604. [31] Brzeziński, Dariusz, and Jerzy Stefanowski. "Accuracy updated ensemble for data streams with concept drift." Hybrid Artificial Intelligent Systems. Springer Berlin Heidelberg, 2011. 155163. [32] Veena M ittal and Indu Kashyap. "Online M ethods of Learning in Occurrence of Concept Drift " International Journal of Computer Applications 117(13):1822, M ay 2015.
How to cite this paper: Veena M ittal, Indu Kashyap , "Empirical Study of Impact of Various Concept Drifts in Data Stream M ining M ethods", International Journal of Intelligent Systems and Applications (IJISA), Vol.8, No.12, pp.6572, 2016. DOI: 10.5815/ijisa.2016.12.08
Authors’ Profiles Veena Tayal received her B.Tech degree in Computer Science from Guru Jambheshwar University of Science & Technology, Hisar in 2007. She received her M . Tech degree in Computer Science from Banasthali University, Rajasthan in 2009. She is a Research Scholar in the Faculty of Engineering at the M anav Rachna International University, Faridabad. She is presently pursuing her doctorate degree in Computer Science under the guidance of Dr. Indu Kashyap. Her research interests include data mining, concept drift, outlier analysis in data streams.
Dr. Indu Kashyap received her Ph.D. in Computer Science from Chaudhary Charan Singh University, M eerut. Dr Indu Kashyap is a Associate Professor in the Computer Science Engineering, Faculty of Engineering & Technology at the M anav Rachna International University, Faridabad . She has guided many M .Tech students in their research work and also guiding Ph.D. students. Her research interests include data mining, mobile computing, computer networks, software engineering . She is an author or coauthor of many research papers in international journals and conferences. She received B.Tech Degree from Jamia Hamdard University, New Delhi in 2004 and M .Tech degree in Computer Science from Banasthali University, Rajasthan in 2006
Copyright © 2016 MECS
I.J. Intelligent Systems and Applications, 2016, 12, 6572
Instructions for Authors Manuscript Submission We invite original, previously unpublished, research papers, review, survey and tutorial papers, application papers, plus case studies, short research notes and letters, on both applied and theoretical aspects. Manuscripts should be written in English. All the papers except survey should ideally not exceed 18,000 words (15 pages) in length. Whenever applicable, submissions must include the following elements: title, authors, affiliations, contacts, abstract, index terms, introduction, main text, conclusions, appendixes, acknowledgement, references, and biographies. Papers should be formatted into A4size (8.27″×11.69″) pages, with main text of 10point Times New Roman, in singlespaced twocolumn format. Figures and tables must be sized as they are to appear in print. Figures should be placed exactly where they are to appear within the text. There is no strict requirement on the format of the manuscripts. However, authors are strongly recommended to follow the format of the final version. Papers should be submitted to the MECS Publisher, Unit B 13/F PRAT COMM’L BLDG, 1719 PRAT AVENUE, TSIMSHATSUI KLN, Hong Kong (Email: ijisa@mecspress.org, Paper Submission System: www.mecspress.org/ijisa/submission.html/), with a cowering email clearly staring the name, address and affiliation of the corresponding author. Paper submissions are accepted only in PDF. Other formats are not acceptable. Each paper will be provided with a unique paper ID for further reference. Authors may suggest 24 reviewers when submitting their works, by providing us with the reviewers’ title, full name and contact information. The editor will decide whether the recommendations will be used or not.
Conference Version Submissions previously published in conference proceedings are eligible for consideration provided that the author informs the Editors at the time of submission and that the submission has undergone substantial revision. In the new submission, authors are required to cite the previous publication and very clearly indicate how the new submission offers substantively novel or different contributions beyond those of the previously published work. The appropriate way to indicate that your paper has been revised substantially is for the new paper to have a new title. Author should supply a copy of the previous version to the Editor, and provide a brief description of the differences between the submitted manuscript and the previous version. If the authors provide a previously published conference submission, Editors will cheek the submission to determine whether there has been sufficient new material added to warrant publication in the Journal. The MECS Publisher’s guidelines are that the submission should contain a significant amount of new material, that is, material that has not been published elsewhere. New results are not required; however, the submission should contain expansions of key ideas, examples, and so on, of the conference submission. The paper submitting to the journal should differ from the previously published material by at least 50 percent.
Review Process Submissions are accepted for review with the same work has been neither submitted to, nor published in, another publication. Concurrent submission to other publications will result in immediate rejection of the submission. All manuscripts will be subject to a well established, fair, unbiased peer review and refereeing procedure, and are considered on the basis of their significance, novelty and usefulness to the Journals readership. The reviewing structure will always ensure the anonymity of the referees. The review output will be one of the following decisions: Accept, Accept with minor revision, Accept with major revision, Reject with a possibility of resubmitting, or Reject. The review process may take approximately three months to be completed. Should authors be requested by the editor to revise the text, the revised version should be submitted within three months for a major revision or one month for a minor revision. Authors who need more time are kindly requested to contact the Editor. The Editor reserves the right to reject a paper if it does not meet the aims and scope of the journal, it is not technically sound, it is not revised satisfactorily, or if it is inadequate in presentation.
Revised and Final Version Submission Revised version should follow the same requirements as for the final version to format the paper, plus a short summary about the modifications authors have made and author’s comments. Authors are requested to the MECS Publisher Journal Style for preparing the final cameraready version. A template in PDF and an MS word template can be downloaded from the web site. Authors are requested to strictly follow the guidelines specified in the templates. Only PDF format is acceptable .The PDF document should be sent as an open file, i.e. without any date protection. Authors should submit their paper electronically through email to the Journal’s submission address. Please always refer to paper ID in the submissions and any further enquiries. Please do not use the Adobe Acrobat PDFWriter to generate the PDF file. Use the Adobe Acrobat Distiller instead, which is contained in the same package as the Acrobat PDFWriter. Make sure that you have used Type 1 or True Type Fonts(cheek with the Acrobat Reader or Acrobat Writer by clicking on File＞Document Properties＞Fonts to see the list of fonts and their type used in the PDF document).
Copyright Submission of your paper to this journal implies that the paper is not under submission for publication elsewhere. Material which has been previously copyrighted, published, or accepted for publication will not be considered for publication in this journal. Submission of a manuscript is interpreted as a statement of certification that no part of the manuscript is under review by any other formal publication. Submitted papers are assumed to contain no proprietary material unprotected by patent or patent application; responsibility for technical content and for protection of proprietary material rests solely with the author(s) and their organizations and is not the responsibility of the MECS Publisher or its editorial staff. The main author is responsible for ensuring that the article has been seen and approved by all the other authors. It is the responsibility of the author to obtain all necessary copyright release permissions for the use of any copyrighted materials in the manuscript prior to the submission. More information about permission request can be found at the web site. Authors are asked to sign a warranty and copyright agreement upon acceptance of their manuscript, before the manuscript can be published. The Copyright Transfer Agreement can be downloaded from the web site. Publication Charges and Reprint No page charges for publications in this journal. Reprints of the paper can be ordered with a price of 150 USD. Electronic: free available on www.mecspress.org.To subscribe, please contact the Journal Subscriptions Department, Email: ijisa@mecspress.org. More information is available on the web site at http://www.mecspress.org/ijisa.
Published on Dec 8, 2016
International Journal of Intelligent Systems and Applications (IJISA) ISSN Print: 2074904X, ISSN Online: 20749058 Volume 8, Number 12, Dec...