Proceedings of The Joint International Conference of

Proceedings of The Joint International Conference of ITCA 2014 and ISCIIA 2014 (ITCA & ISCIIA 2014)

Edited by Yong HE, Jie CHEN, Jinhua SHE, Xin CHEN

Changsha, China, 15-20 September, 2014

School of Information Science and Engineering, Central South University

The Joint International Conference of ITCA 2014 and ISCIIA 2014 Changsha, China, 15-20 September, 2014 Organizers International Fuzzy System Association Central South University, China Beijing Institution of Technology, China Tokyo University of Technology, Japan Editors: HE Yong CHEN Jie SHE Jinhua CHEN Xin Honorary Chair: HIROTA Kaotu, Japan Organizing Committee: Chairperson: HE Yong, China Co-Chairperson: PENG Zhihong, China SHE Jinhua, Japan CHEN Xin, China International Program Committee: Chairperson: WU Min, China Co-Chairperson: CHEN Jie, China OHYAMA Yasuhiro, Japan Members: ULLAH AMM Sharif, Japan BARGIELA Andrzej, UK BEDE Barnabas, USA CAO Weihua, China CHEN Deyun, China CHEN Jianer, USA CHEN Shumei, Japan CHEN Xin, China CHEN Zhigang, China CLIFTON A. David, UK DAI Yaping, China DAN Jingpei, China DAN Ralescu, USA DENG Fang, China DONG Fangyan, Japan FANG Hao, China

HE Yong, China HONDA Katsuhiro, Japan HUANG Dongjun, China ILIYASU M. Abdullah, Japan LIU Guoping, UK LIU Xiangdong, China LUO Zhongming, China MITSUHASHI Kaoru, Japan NGO Long Thanh, Vietnam OHNISHI Shin-ichi, Japan OKAMOTO Kazushi, Japan PEDRYCZ Witold, Canada PENG Hui, China PENG Jun, China PENG Zhihong, China SUKUMA Yuji, Japan SHE Jinhua, Japan SIRADJUDDIN Indah Agustien, Indonesia SU Wei, China SUN Jian, China TAKAMA Yasufumi, Japan WANG Guojun, China WANG Qinglin, China XIE Fuding, China XIN Bin, China XIN Xin, Japan XU Li, Japan YAMAZAKI Yoichi, Japan YAN Fei, China YI Jiangqiang, China ZHU Guohun, China International AdvisoryCommittee: LEE KwangHyung, Korea KAMEDA Hiroyuki, Japan KAWATA Seiichi, Japan KIM Dong Hwa, Korea JIN Jianping, China NAKAMURA Taichi, Japan OHNO Sumio, Japan

Proceedings of the Joint International Conference of ITCA 2014 and ISCIIA 2014 (ITCA & ISCIIA 2014) Changsha, China, 15-20 September, 2014

Edited by Yong HE, Jie CHEN, Jinhua SHE, Xin CHEN

Organized by

CONTENTS Plenary Lecture Fuzzy Set Representation of KANSEI Texture for Online Shopping Hidenori SAKANIWA, Fanqyan DONG, Kaoru HIROTA

1

Fuzzy Transfer Learning for Prediction Jie LU

8

Modeling of High Temperature Gas Flow 3D Distribution in BF Throat Based on the Computational Fluid Dynamics 9 Jianqi AN, Kai PENG, Weihua CAO, Min WU, Yang LIU

Session 1: Carbon Efficiency Improvement & Advanced Control A Novel Method for Recasting an n-D Fornasini-marchesini Model into an n-D Roesser Model 17 Nozomu SATO, Li XU, Hua CHENG, Shin-ya MATSUSHITA, Hirokazu MADOKORO, Nobuhiro SHIMOI Compensation and Optimisation for Aperiodic Disturbances and Input Dead Zone in Repetitive Control System Jieqiong LIU, Min WU, Fang GAO, Yong HE

23

A Hierarchical Experimental Simulation Platform of Coking Production Qi LEI, Hui YAN, Min WU

29

Session 2: Intelligent Systems & Computational Intelligence 1 A Command Based Real-Time Navigataion System for Overseas Visitors Supported by AR 37 Keqiang HUANG, Aratsusi KIMURA, Osamu OKUMURA, Azusa MIYASE, Seiichi KAWATA, Junfu CHEN Distributed Cooperation Based Priority Coverage Control Strategy for Mobile Sensors Zhi ZHENG, Zhihong PENG

45

Design and Implement of Intelligent Human-Computer Interface on Vehicles based on Event-driven Jia ZHANG, Shengli XU, Fang DENG

52

Session 3: Intelligent Systems & Computational Intelligence 2 Neural Network Size Estimation Method based-on Hierarchical Force-Directed Graph Drawing for Multi-task Learning 59 Atsushi SHIBATA, Fanqyan DONG, Kaoru HIROTA Measuring Polarization with Median Filtering based on Non-Orthogonal Angles Liyan PANG, Yange LV, Guohun ZHU

65

Ischemia Diagnosis using Fuzzy Association Rule Mining on ECG Signal Tianyu LI, Fangyan DONG, Kaoru HIROTA

69

Locating Informative Bright Region in Tunnel Scenes using Lighting and Traffic Lane Cues Jiajun LU, Fangyan DONG, Kaoru HIROTA

75

I

Study of Compressive Channel Estimation in MIMO-OFDM Two-Way Relay Networks Guan GUI, Hongyun WEI, Li XU

83

Guidance Law of Multiple Missiles Engaged in Cooperative Salvo Attack Jie ZENG, Lihua DOU, Bin XIN

90

Session 4: Advanced Control Theory & Applications An Elementary Operation Approach to the Realization of Multidimensional Systems in Fornasini-Marchesini State-space Model: the MIMO case 95 Kohei YAMAUCHI, Shi YAN, Guan GUI, Shinya MATSUSHITA, Li XU Quantitative Analysis of RBAC Model with N-dimensional Security Entropy Qian HE, Yaping DAI, Linhui ZHAO, Likun CAI

100

PSO Based Deterministic ESN Models for Stock Price Forecasting Jingpei DAN, Wenbo GUO, Weiren SHI, Bin FANG, Tingping ZHANG

104

Classification of Corresponding Points for 3D Measurement Using Moving Monocular Camera attached with 6-axis Sensor 110 Toshihiro AKAMATSU, Fangyan DONG, Kaoru HIROTA Robust H∞Damping Control of Multi-FACTS Devices for Stability Enhancement of Power Systems with Signal’s Time Delay 117 Fang LIU, Min WU, Danyun LI, Yong HE, Ryuichi YOKOYAMA Novel Realization of Adaptive Sparse Sensing using Reweighted Zero-Attracting Least Mean Forth Algorithm 124 Guan GUI, Shinya MATSUSHITA, Li XU

Session 5: Sensor & MEMS Practice of Active Learning Based on Craftsmanship Liang XIAO, Zhejun FANG, Jie ZHANG, Jinhua SHE, Yasuhiro OHYAMA

130

The Energy and Transmission Optimization based on Dynamic Spectrum Sensing in Cognitive Radio Sensor Networks 134 Yi LI, Jun PENG, Weirong LIU, Kaiyang LIU and Fu JIANG Buffer Constraints Aware Data Gathering with Emergencies for Wireless Sensor and Actuator Networks Weirong LIU, Yun HE, Shuo LI, Fu JIANG, Jun PENG

142

Fast Differential Evolution Algorithm based Energy-Efficient Optimization for Cooperative Spectrum Sensing in Cognitive Radio Sensor Network 149 Weirong LIU, Gaorong QIN, Fu JIANG, Shuo LI, Jian HE, Jun PENG Distributed Q Learning based Energy-Efficiency Optimization with Spectrum Decision for in Cognitive Radio Sensor Network 157 Jian HE, Jun PENG, Fu JIANG, Gaorong QIN, Weirong LIU II

Hypergraph Based Data Allocation for Online Social Networks Wenyin YANG, Guojun WANG

163

Session 6: Robot Systems & Control Ground Target Tracking and Collision Avoidance for UAV Based Guidance Vector Field Zhimin CHEN, Zhihong PENG

173

An Improved Kalman-filter for Visual Servoing Applied in Nonholonomic Robot with the FOV Constraint 180 Yuan FANG, Zhiwu HIANG, Yabo WANG, Wentao YU, Xiaoyong ZHANG Information-Driven Multi-Robot Behavior Adaptation to Intention in Human-Robot Interaction Luefeng CHEN, Zhentao LIU, Min WU, Fangyan DONG, Kaoru HIROTA

185

Representation of Quantum Emotion Space using Bloch Sphere Fei YAN, Abdullah M. ILIYASU, Zhentao LIU, Zhengang JIANG, Fangyan DONG, Kaoru HIROTA

195

Motion Control of a Mobile Wheeled Inverted Pendulum Using Equivalent Input Disturbance Approach 201 Qi SHI, Zhejun FANG, Jinhua SHE, Junya IMANI, Yasuhiro OHYAMA Development of Electric Walker Kosuke YAMAGUCHI, Tomo ISHIKAWA, Kaoru MITUHASI, Yasuhiro OHYAMA

III

208

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Fuzzy Set Representation of KANSEI Texture for Online Shopping Hidenori SAKANIWA*,**, Fanqyan DONG*, Kaoru HIROTA* * Department of Computational Intelligence and Systems Science, Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology Yokohama, Kanagawa 226-8502, Japan ** Yokohama Research Laboratory, Hitachi Ltd., Yokohama, Kanagawa 244-0817, Japan

Abstract

visualization method of these 2 vectors is also proposed. The standard deviation vector in [0,1]3 provides the individual difference degree of Kansei Texture in "PuruPuru - GotsuGotsu", "KachiKachi - FuwaFuwa", and "ButsuButsu - PikaPika" axes. If the standard deviation vector is big, then buyers may request sample sending to confirm fitness of the good and sellers may make a strategy for return-goods. If the standard deviation vector is small, then buyers may make a decision smoothly and sellers may concentrate on other goods. The standard deviation vector information is utilized to operate online shopping effectively. The standard deviation vector with the average vector of Kansei Texture is illustrated and is added to the virtual online shopping web site. A subjective evaluation experiment is done for 17 subjects and 20 goods by showing the proposed virtual online shopping web site, and the positive effect of individual difference to buyers/sellers is confirmed. In 2, fuzzy set representation of Kansei Texture is proposed. A visualization method of Kansei Texture fuzzy set representation is proposed in 3. In 4, a subjective evaluation experiment is done to show the availability of individual difference of Kansei Texture.

A fuzzy set representation method of Kansei Texture is proposed to express individual difference of Kansei Texture feelings for the purpose of mainly online shopping. It provides buyers the criteria if sample sending request is necessary or not according to the variance degree of individual difference and it also offers sellers the possibility information of return-goods in the case of big individual difference. The correlation coefficient of the degree of individual difference and sample demand is 0.78(P 0 in contrast to randomly generated input weights in standard ESN. The sign of v is determined randomly by a random draw from Bernoulli distribution of mean 1/2. Generally, the deterministic form of ESNs has the following properties, which simplifies the ESN construction by only setting two free parameters and enables a more thorough theoretical analysis of the ESN performance [10]:

105

1) a simple fixed non-random reservoir topology with full connectivity from inputs to the reservoir , 2) a single fixed absolute weight value r for all reservoir connections and 3) a single weight value v for input connections, with aperiodic pattern of input signs.

3. STOCK PRICE FORECASTING BASED ON THE PROPOSED PSO-DESN MODELS In this section, we present the proposed PSO-DESN model and its application for stock price prediction. To verify performances of PSO-DESN models for stock price prediction, three types of reservoirs, DLR, DLRB, and SCR are compared to standard ESN model under different conditions. Firstly, the comparative experiments are under the same input weight Win without applying PSO, then the forecasting performances and running times of those models are compared with parameters optimized by PSO.

2.3 Particle Swarm Optimization Particle swarm optimization is a popular computational technique developed by Kennedy and Eberhart [13] that is based on the social behavior of birds flocking to look for food. Sierra and Coello [14] note two reasons for PSO’s popularity. First, since it is relatively simple, its implementation is straightforward; and second, it has been found to be very effective in a variety of applications, producing very good results at very low computational cost. PSO has been found to be effective in optimization problems requiring realvalued decision variables [15, 16]. Hassan et al. [17] report that while PSO’s performance is comparable to GA, PSO is computationally more efficient than GA. In a particle swarm optimizer, each particle adjusts its flying according to its own flying experience and its companions’ flying experience. Each individual is named as a “particle” which, in fact, represents a potential solution to a problem. Each particle is treated as a point in an N dimensional space. The ith particle is represented as X i  xi1 , xi 2 ,..., xiN  . The best previous position (the position giving the best fitness value) of any particle is recorded and represented as Pi   pi1 , pi 2 ,..., piN  . The index of the best particle among all the particles in the population is represented by the symbol g. The rate of the position change (velocity) for particle i is represented as Vi  vi1 , vi 2 ,..., viN  . The particles are manipulated according to the following equation: vid  vid  c1  rand     pid  xid   (3) c2  Rand    p gd  xid



3.1 The Proposed PSO-DESN Models For each ESN model (ESN, DLR, DLRB, SCR), parameter vector (Spectral Radius, V, Range, connectivity) (Spectral, Radius, V, R) (Spectral, Radius, V, R, b) (Spectral, Radius, V, R) is set up, respectively. As shown in Fig.5, the position and velocity in each particle are setup according to the range of each parameter initially, then the ESN models are constructed based on the parameters that each particle has. The models are trained and tested by using training dataset and testing dataset respectively. The optimal particle is found out by the fitness value, and the optimal model then is found by the particle swarm optimization algorithm for every iteration. In this study, Euclidean distance is used to determine the neighborhood, i.e., search scope and only search the optimal value in the neighborhood for each particle. With the iteration of searching, the search scope is eventually expanded to the whole particle swarm. 3.2 Experimental Data Since S&P500 is usually considered as the benchmark for the United States’ equity performance, it is popular for stock price prediction research. The daily trading price data from 1st June, 2010 to 20th February, 2014 (totally 930 days) of 374 stocks in S&P500 is adopted in the experiments. All the data come from Yahoo Finance. For each stock price data in the experiments, 80% data points are used as training set, and the remaining 20% are used for testing.



xid  xid  vid

(4) where c1 and c2 are two positive constants, rand() and Rand() are two random functions in the range [0,1]. The second part of the equation (3) is the “cognition” part, which represents the private thinking of the particle itself. The third part is the “social” part, which represents the collaboration among the particles [13]. The equation (3) is used to calculate the particle’s new velocity according to its previous velocity and the distances of its current position from its own best experience (position) and the group’s best experience. Then the particle flies toward a new position according to equation (4). The performance of each particle is measured according to a predefined fitness function, which is related to the problem to be solved.

3.3 Data Preparation In most time series prediction researches, the original data are usually normalized in order to make the preprocessed data in a desired range. In this study, all the original data are normalized as follows: X  k  x  min x  maxx   min x   t (5) where x is the original data, min(x) is the minimum data, max (x) is the maximum value in the data, while k and t are two coefficients to make the preprocessed data X in a desired range.

106

MSE 

Initialize particles (Position and velocity)

1 n

n

2

 ( y t  x t)

(6)

t 1

where xt denotes a given time series, t = 1,…, n, n∈N, yt denotes the forecast of value at time t. 3.5 Experimental Results Construct an ESN model Compute the fitness value for each particle

In the experiment, each ESN model (standard ESN, DLR, DLRB, SCR) is built with 96 internal units. For each stock in this dataset, the next trading day closing price is predicted by its previous 30 closing prices. The average forecasting accuracy for each model is achieved by 1000 times tests.

Train the ESN model using training dataset  Test the ESN model using testing dataset

3.5.1 Deterministic ESN models The forecasting performances of these ESN models are compared under the condition of same Win value. In this experiment, all the values in the Win, v=0.5, and the sign of v is the probability of 50% randomly determined. In DLR, DLRB, and SCR, all the values in internal weight matrix W, r= 0.5, the feedback connection values in W of DLRB, b=0.05. The experimental results are shown in table 1 partly which indicate the deterministic ESN models perform competitively in forecasting accuracy compared to the standard ESN model under the condition of same constructions.

Find out the best accuracy of the ESN model and the best position of the particles

Search for gbest：the position of the particle giving the best fitness value

3.5.2 PSO-DESN models The number of particles is set to be 20 and the initial iteration number is 50. In order to reduce the influence of the convergence speed, the number of experiments increases adaptively during the process of searching for the optimal parameters in the particle, that is, when the optimal parameters of two successive rounds of experiments are equal, the experiment will be executed one more time. Table 2 shows the experimental results for the same stock data as in table 1. The average prediction performance of each forecasting model is greatly improved by 63.80%, 64.25%, 63.79%, and 64.27%, respectively by applying PSO algorithm, which indicates PSO-DESN models outperform the deterministic ESN and standard ESN models. Meanwhile the PSO-DESN models have obvious advantages in efficiency which is indicated by the average time saving of 32.60%, 32.49%, and 32.70% for DLR, DLRB, and SCR compared to the standard ESN model, respectively. The experimental results indicate that the proposed PSO-DESN models improve in forecasting accuracy compared to the general deterministic ESN models by applying PSO algorithm to optimize the parameters of the reservoir structures, while running time are increased accordingly. The PSO-DESN models, however, improve in efficiency compared to the standard ESN models.

Compute the new velocity according to the current velocity and gbest for each particle

Update the position according to the new velocity for each particle

NO

Iteration Finish

YES

Output the best forecasting model and its performance

Fig.5 The Proposed PSO-DESN model

3.4 Evaluation Measure A commonly used measure, Mean Squared Error (MSE) as shown in Equation (6), is utilized to evaluate forecasting accuracy.

107

Table 1. Comparisons of Average forecasting accuracy (MSE) Average forecasting accuracy (r=0.5 b=0.05 v=±0.5)

Stocks ESN

DLR

DLRB

SCR

'A'

3.814025

2.947236

2.87385

3.042726

'ALL'

0.729146

0.737923

0.71832

0.730556

'BAX'

0.47736

0.504663

0.504453

0.496313

'CAT'

1.227007

1.21849

1.221667

1.221545

'COH'

0.702417

0.670763

0.673214

0.67075

'DF'

0.23828

0.262445

0.262805

0.264985

'IFF'

2.963587

2.798408

2.70644

2.977701

'OI'

0.558538

0.476727

0.480012

0.47534

'PKI'

3.549767

2.748976

2.64719

2.657761 6.409925

'RL'

6.35157

6.377153

6.40885

'SRE'

7.084271

6.978886

6.78874

7.210348

'VAR'

1.697554

1.58872

1.617942

1.613488

'WY'

0.15677

0.16673

0.165503

0.166299

'WYN'

1.538461

1.378661

1.35106

1.37561

'XOM'

1.870293

1.827589

1.824464

1.80176

'YUM'

2.818102

2.876777

2.806851

2.79861

Table 2. Comparisons of optimal forecasting accuracy (MSE) and running time (s) MSE Stock

Running Time(s)

ESN

DLR

DLRB

SCR

ESN

DLR

DLRB

SCR

'A'

0.971271

0.993583

0.976638

0.965038

3.0548

2.0047

2.0111

1.9950

'ALL'

0.316965

0.339634

0.344472

0.341823

3.0333

2.0152

2.0053

2.0182

'BAX'

0.340443

0.312214

0.321759

0.330580

3.1426

2.0273

2.0160

2.0144

'CAT'

1.027453

1.051898

1.010183

1.024609

3.1897

2.0966

2.1136

2.0888

'COH'

0.494817

0.472020

0.499416

0.472057

3.1874

2.0405

2.0684

2.0497

'DF'

0.127926

0.130516

0.111625

0.122880

3.0261

2.0344

2.0326

2.0434

'IFF'

0.722405

0.650138

0.769070

0.688414

3.1439

2.3356

2.3050

2.1724

'OI'

0.206284

0.202518

0.194742

0.182031

3.0751

2.0934

2.1135

2.0954

'PKI'

0.226959

0.238861

0.214321

0.233244

3.1460

2.2611

2.2251

2.2273

'RL'

5.372283

5.430342

5.388721

5.322740

3.0202

2.0811

2.0819

2.0820

'SRE'

0.619618

0.605086

0.609867

0.609244

2.9712

2.0678

2.0860

2.0709

'VAR'

0.705757

0.693458

0.632998

0.684081

3.2103

2.1057

2.1053

2.1026

'WY'

0.134003

0.128507

0.126444

0.132080

3.0581

2.0961

2.1055

2.0923

'WYN'

0.737564

0.722017

0.744484

0.786635

3.1492

2.1058

2.1085

2.1094

'XOM'

0.994823

0.973868

0.909183

0.920857

3.2169

2.1102

2.1081

2.1101

'YUM'

1.261219

1.219010

1.304283

1.237525

3.1132

2.1556

2.1042

2.0971

types) have been applied to predict stock price. The forecasting performances have been investigated on the benchmark dataset (S&P500). The experimental results on on S&P 500 dataset show that the deterministic ESNs have competitive forecasting performances as the standard ESN model, while the proposed PSO-DESN models have improvement in average accuracy by about 63.80%, 64.25%, 63.79%, and 64.27%, respectively, which indicates PSO-DESN models outperform the deterministic ESN and standard ESN models without applying PSO algorithm. Meanwhile the PSO-DESN models have obvious advantages in efficiency which is

4. CONCLUSION The deterministic ESN model optimized by PSO (PSO-DESN) is proposed to improve time series forecasting and applied to stock price forecasting. In standard ESNs, the reservoirs are random constructed, which makes researchers and practitioners have to rely on trials and errors. Differently, the deterministic ESNs have their reservoirs constructed deterministically so that the structures of ESN models are simple and easy to be applied. Therefore, deterministic ESNs (three typical

108

indicated by the average time saving of 32.60%, 32.49%, and 32.70% for DLR, DLRB, and SCR compared to the standard ESN model while insignificant improvement in forecasting accuracy. Since PSO-DESNs have advantages of higher forecasting accuracy and efficiency as indicated by the experimental results, they have great prospects in applications of stock price prediction and even other time series forecasting applications.

[9]

[10]

[11]

ACKNOWLEDGMENTS The research work is partially supported by the Fundamental Research Funds for the Central Universities in China (CDJZR12180004) and the Major State Basic Research Development Program of China (973 Program)(No. 2013CB329100)

[12]

References

[13]

[1] H. Jaeger, H. Haas, “Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication”, Science, Vol.304, No.5667, 2004, pp.78-80. [2] Q. Song, X. Liu, and X. Zhao, “Short-term Traffic Flow and Hourly Electric Load Forecasting Algorithm based on Echo State Neural Networks”, International Journal of Digital Content Technology & its Applications, Vol. 6, No.4, 2012. [3] Zhiwei Shi and Min Han, “Support vector echo-state machine for chaotic time-series prediction”, IEEE Transactions on Neural Networks, Vol.18, No.2, 2007, pp.359-372. [4] C. Sheng, J. Zhao, Y. Liu, W. Wang. Prediction for noisy nonlinear time series by echo state network based on dual estimation. Neurocomputing, Vol. 82, 2012, pp.186-195. [5] Kara, Yakup, Melek Acar Boyacioglu, and Ömer Kaan Baykan, “Predicting direction of stock price index movement using artificial neural networks and support vector machines: The sample of the Istanbul Stock Exchange”, Expert systems with Applications, Vol.38, No.5, 2011, pp.5311-5319. [6] F. Zhai, X. Lin, Z. Yang, Y. Song, “Financial time series prediction based on echo state network”, In Natural Computation (ICNC), 2010 Sixth International Conference on, IEEE, Vol. 8, pp. 3983-3987. [7] H. Zhang, J. Liang, Z. Chai, “Stock Prediction Based on Phase Space Reconstruction and Echo State Networks”, Journal of Algorithms & Computational Technology, Vol.7, No.1, 2013, pp.87-100. [8] X. Lin, Z. Yang, Y. Song, “Short-term stock price prediction based on echo state networks”, Expert systems with applications, Vol.36, No.3, 2009,

[14]

[15]

[16] [17]

109

pp.7313-7317. X. Lin, Z. Yang, Y. Song, “ Intelligent stock trading system based on improved technical analysis and Echo State Network”, Expert systems with Applications, Vol.38, No.9, 2011, pp.11347-11354. A. Rodan and P. Tiňo, “Minimum complexity echo state network”, IEEE Trans on Neural Networks, Vol.22, No.1, 2011, pp.131-144. A. Rodan and P. Tiňo, “Simple deterministically constructed cycle reservoirs with regular jumps”, Neural computation, Vol.24, No.7, 2012, pp.1822-1852. A. Rodan and P. Tiňo, “A. Short term memory in input-driven linear dynamical systems”, Neurocomputing, Vol.112, 2013, pp.58-63. J. Kennedy, R. Eberhart, “Particle swarm optimization”, Proceedings of the 1995 IEEE International Conference on Neural Networks, Piscataway, New Jersey, IEEE Service Center, 1995. M.R. Sierra, C.C. Coello, “Multi-objective particle swarm optimizers: a survey of the state-of-the-art”, International Journal of Computational Intelligence Research, 23, 2006, pp.287–308. M. Clerc, J. Kennedy, “The particle swarm explosion, stability, and convergence in a multidimensional complex space”, IEEE Trans. on Evolutionary Computation,Vol. 6, No.1, 2002, pp.58–73. A. Engelbrecht, Fundamentals of Computational swarm Intelligence, John Wiley & Sons, 2005. R. Hassan, B. Cohanim, O. de Weck, G. Venter, A comparison of particle swarm optimization and the genetic algorithm, in: Proceedings of the 46th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, No. AIAA-2005-1897, Austin, TX, 2005.

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Classification of Corresponding Points for 3D Measurement Using Moving Monocular Camera attached with 6-axis Sensor Toshihiro AKAMATSU*, Fangyan DONG*, Kaoru HIROTA* * Department of Computational Intelligence and Systems Science, Tokyo Institute of Technology G3-49, 4259 Nagatsuta, Midori-ku, Yokohama-city 226-8502, Japan Abstract

attached with moving monocular camera is proposed. In addition, it is used to identify the fundamental matrix, and the foundation of 3D measurement method is established. The proposed method classifies corresponding points into two categories (i.e., belonging to still/moving objects) with high accuracy by using two frames taken by moving monocular camera in the situation that both still and moving objects exist. Furthermore, it is possible to pick out more efficient corresponding points to identify the fundamental matrix. The corresponding points classification experiments are performed on a laptop PC using original CG images which include still and moving objects. The classification method of corresponding points using the device which is composed of a 6-axis sensor to a moving monocular camera is proposed in 2. The results of corresponding points classification experiments are shown in 3.

A classification method of corresponding points is proposed, which uses a moving monocular camera attached with a 6-axis sensor. It classifies corresponding points between two consecutive frames containing still/moving objects and chooses appropriate corresponding points for 3D measurement. Corresponding point classification experiment with original CG images shows that accuracy, precision, and recall are 0.94, 0.85, and 1.00, respectively. In addition, it is confirmed that erroneous correspondences are removed by outlier detection using Interquartile Range for proposed evaluation function calculated on each corresponding point. The proposed method is planning to be included in 3D measurement method with actual images containing still/moving objects and also to be applied to obstacles avoidance for vehicles or vision system for mobile robots.

2. CORRESPONDING POINTS CLASSIFICATION METHOD USING MOVING MONOCULAR CAMERA AND 6-AXIS SENSOR

Keywords: 3D Measurement, Corresponding Points Classification, Fundamental Matrix Estimation, 6-axis Sensor, Moving Camera.

1. INTRODUCTION

2.1 Classification Condition of Corresponding Points The relation between corresponding points that belong to still objects is formulated as fundamental equation, and it is represented as

To measure depth to a still/moving object and its shape using a still stereo camera system, it is general that corresponding points are extracted from two images containing the object, and that fundamental matrix is calculated using fundamental equation [1-3]. At least seven pairs of corresponding points are necessary to obtain the fundamental matrix [4]. Furthermore, considering practical applications, for example in-vehicle cameras for obstacle avoidance and mobile robot’s vision, two images are necessary per frame, which takes much computational cost. On the other hand, moving monocular camera system needs only one image per frame [5], but it is difficult to choose seven pairs of corresponding points that belong to still objects in a real situation as still and moving objects are mixed. In the case of either, there exist problems such as erroneous correspondence when corresponding points are extracted [6] and computational cost caused by bundle adjustment [7]. The corresponding points classification method using 6axis sensor (3-axis accelerometer and 3-axis gyro sensor)

′

0,

1

where and ′ are image coordinates of extracted corresponding points, and they are expressed in homogeneous coordinates. The origin of image coordinate is defined as center of image, x-axis and y-axis are defined as upward and rightward of image, respectively. In the situation in which still and moving objects are mixed, classification condition between them is defined as and ⇒ ′ 0 , (2) and ⇒ ′ 0 using Eq.(1). 2.2 Proposal of Classification Criteria

110

belong to still objects is that , namely,

Using 6-axis Sensor It is necessary to calculate the fundamental matrix for 3D measurement. In the situation that still and moving objects are observed by moving monocular camera, however, it is difficult to choose seven pairs of corresponding point and calculate the fundamental matrix. Accordingly, a method that classifies each corresponding point into still or moving objects is proposed using a device composed of a 6-axis sensor to a movement monocular camera. The fundamental matrix can be decomposed into intrinsic parameters and extrinsic parameters of the camera, and it is represented as ,

′

0 0

Under a certain condition in which erroneous correspondence happens, or moving objects moves parallel to the device and so on, corresponding points not to belong to still objects may satisfy Eq.(11). So, corresponding points chosen by Eq.(11) are only candidates. Accordingly, a method to extract corresponding points to belong to still objects more likely from the candidates is proposed. Corresponding points on still objects have same fundamental matrix in the fundamental equation, so the solutions of Eq.(6), namely, is also the same. Consequently, the evaluation function is defined as ,

0 0 0

0

0 0 1

0 0 0

0

0 0 1

0.

(5)

In Eq.(5), and can be obtain from 6-axis sensor, so it can return to the quadratic equation about . Consequently, the condition for corresponding points to belong to still objects is that Eq.(5) has solutions with different signs. The left hand side of Eq.(5) is developed and coefficient of each term is denoted with A, B, and C, respectively. Then, Eq.(5) is represented by 0.

3. CORRESPONDING POINTS CLASSIFICATION EXPERIMENTS 3.1 Environment of Corresponding Points Classification Experiments

(6)

Simulation experiments are performed to verify the proposed corresponding points classification method. Two CG images created by Blender [8] are used as input images. They represent images before and after the movement of the device, respectively. Corresponding points are extracted from them using Oriented-BRIEF (ORB) [9]. They are classified in ones to belong to still objects or not, and classification accuracy is confirmed. All experiments are performed on a laptop with Windows 7 Professional 64 bits OS, Intel® Core™ i5-3210M 2.50 GHz CPU, and 8.00 GB RAM. The system is coded by Python 2.7.

In addition, putting that , , ,1 , ′ are expressed as

(7)

′, ′, 1 , coefficients ′ ′ .

′

,

,

, and

,

(8)

(12)

following axis of Eq.(6), values of corresponding points that do not to belong to still objects are detected for outliers, so they can be removed. The value of is used as parameter which represents accuracy of each corresponding point. Corresponding points in the range, where their values of are high-frequency, are considered to be accurate corresponding points. Accordingly, the fundamental matrix is possible to be estimated with higher accuracy and lower computational cost by using them.

(4)

Substituting Eq. (4) for Eq. (1), it is expressed as ′

(11)

2.3 Removal of Erroneous Correspondence and Accuracy Evaluation of Corresponding Points

(3)

0 0 . 1

have different sign

0.

where and ′ are intrinsic parameters, is rotation matrix, is translation vector, ∙ is operator to express skew symmetric matrix. Assuming that aspect ratio is equal to 1, skew coefficient is equal to 0, image center locates at 0, 0 , focal length is constant before and after the movement of the device, and ′ are expressed as 0 0

and

(9)

3.2 Simulation Experiments of Corresponding Points Classification

(10)

Therefore, the condition for corresponding points to

111

Fig.2 The image from viewpoint after the device movement.

Fig.1 The image from viewpoint before the device movement.

Fig.3 The bird’s-eye view image of positional relation between objects and device.

Fig.4 The image of extracted corresponding points.

corresponding points are extracted, and ones to belong to the cube (still object) are 17 pairs. First, pairs of points to belong to still objects or not are classified from 52 points using Eq.(11). Classification results are shown in Table 1 and the image of corresponding points classified to belong to still objects is shown in Fig.5. Pairs of corresponding points classified to belong to still objects are 20 pairs, but 3 pairs of erroneous correspondence are included in them. In addition, the results summarized in indexes are shown in Table 2. Next, value of is calculated for each 20 pairs. Pairs of corresponding points that are more likely to belong to the cube are chosen to remove outliers in values of . Fig.6 is a histogram of . It is confirmed that values of erroneous correspondence are outliers. The

Input images are shown in Fig.1 and Fig.2. Fig.1 is an image before the device moves and Fig.2 is one after the device moves. A cube is a still object, and a car and a helicopter are moving objects in the input images. The car runs parallel on the floor, and the helicopter travels in the air. Moving directions and positional relations of all objects and the device are shown in Fig.3. Assuming that x-axis, y-axis, and z-axis are defined as upward, rightward, and frontward for the device, respectively, translation vector of the device is expressed as is ⁄12 0.50 1.00 1.00 , and rotation angle rad around x-axis. Corresponding points extracted from input images using ORB are shown in Fig.4. In detail, 52 pairs of

112

Fig.5 The image of classified corresponding points that belong to still objects. Table 3: The experimental result of corresponding points classification.

True Positive False Positive False Negative True Negative Total

Fig.6 The histogram of

image after removing outliers is shown in Fig.7. Interquartile Range (IQR) is used for removal of outliers. Comparing to Fig.5, it is confirmed that 3 pairs of erroneous correspondence are removed and the classification is perfectly performed (Table 3).

for each candidate of still

corresponding point. Table 1: The experimental result of corresponding points classification.

True Positive False Positive False Negative True Negative Total

3.3 Validation of Robustness for Noise of 6-axis Sensor

17 3 0 32 52

Considering actual performance of a 6-axis sensor, it is general that noises are added to the sensor data. Accordingly, classification performance is validated in the case of adding noises to translation vector and rotation angle. Noises are added following a normal distribution whose average is 0.0 and variance is . Each translation vector and rotation angle added noise are shown in Table 4. Classification results when 0.1, 0.2, 0.3, 0.4, and 0.5 are shown in Table 5, 6, 7, 8, and 9, histograms are shown in Fig.8, 9, 10, 11, and 12, and images of corresponding points classified to belong to still objects are shown Fig.13, 14, 15, 16, and 17. Only pairs of

Table 2: The index of classification experimental result.

Accuracy Precision Recall F-measure

17 0 0 35 52

0.94 0.85 1.00 0.92

Fig.7 The image to remove erroneous correspondence.

113

Table 4. Translation vectors and rotation angles added noises 0.1 0.46 0.95 1.10 Average error of Error of

0.2 0.42 1.21 1.11

0.5 0.72 0.69 1.27

0.08

0.16

0.22

0.32

0.34

0.22

0.28

0.21

0.37

0.11

0.16

0.07

0.20

0.41

0.1.

17 0 0 35 52

Table 6: The corresponding points classification result when

True Positive False Positive False Negative True Negative Total

0.2.

16 0 1 35 52

Table 7: The corresponding points classification result when

True Positive False Positive False Negative True Negative Total

Fig.9 histogram of

when

0.2 point.

0.3.

17 0 0 35 52

Table 8: The corresponding points classification result when

True Positive False Positive False Negative True Negative Total

0.4.

17 3 0 32 52

Table 9: The corresponding points classification result when

True Positive False Positive False Negative True Negative Total

Fig.8 histogram of

0.4 0.76 0.82 1.33

0.29

Table 5: The corresponding points classification result when

True Positive False Positive False Negative True Negative Total

0.3 0.71 0.91 0.85

when

0.3 point.

Fig.11 histogram of

when

0.4 point.

Fig.12 histogram of

when

0.5 point.

0.5.

11 3 6 32 52

when

Fig.10 histogram of

0.1 point.

114

Fig.13 Corresponding points classified to belong to still objects when

0.1.

Fig.14 Corresponding points classified to belong to still objects when

0.2.

Fig.15 Corresponding points classified to belong to still objects when

0.3.

Fig.16 Corresponding points classified to belong to still objects when

115

0.4.

Fig.17 Corresponding points classified to belong to still objects when

0.5.

Vision”, Vol. 1, 2002, pp. 338-343. [2] R. Hartley and A. Zisserman, “Multiple View Geometry in Computer Vision Second Edition”, Cambridge University Press, March, 2004. [3] K. Kanatani and Y. Sugaya, “Compact fundamental matrix computation”, IPSJ Transactions on Computer Vision and Applications, Vol. 2, 2010, pp. 59-70 [4] O. D. Faugeras, “Stratification of threedimensional vision: projective, affine, and metric representations”, Journal of the Optical Society of America, Vol. 12, No. 3, 1995, pp. 465-484. [5] Z. Hu and Z. Tan, “Depth recovery and affine reconstruction under camera pure translation, Pattern Recognition”, Vol. 40, Issue 10, 2007, pp. 2826-2836. [6] Z. Zhang, R. Deriche, O. D. Faugeras, and Q. Luong, “A Robust Technique for Matching Two Uncalibrated Images Through the Recovery of the Unknown Epipolar Geometry”, ReserchReport 2273, 1994. [7] K. Kanatani and Y. Sugaya, “Implementation and evaluation of bundle adjustment for 3-D reconstruction”, Proceedings of the 17th Symposium on Sensing via Imaging Information, 2011, pp. IS402-1-IS4-02-8. [8] http://www.blender.org/ [9] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, “ORB: an efficient alternative to SIFT or SURF”, Proceedings of the IEEE International Conference on Computer Vision, 2011, pp. 2564-2571.

corresponding points to belong to the cube are extracted when 0.1 , 0.2, and 0.3, but erroneous correspondences are extracted when 0.4 and 0.5 from this results. That is to say, it is confirmed that the proposed method extracted only pairs of corresponding points to belong to still objects, even if there are around 20% of sensor noises

4. CONCLUSION Corresponding points classification experiments are performed using CG input images in the situation that still and moving objects are mixed. Pairs of corresponding points to belong to still objects are 17 pairs in all extracted corresponding points by ORB, and 20 pairs are classified to belong to still objects by proposal. As the result, accuracy, precision, recall, and F-measure are 0.94, 0.85, 1.00, and 0.92, respectively. In addition, it is confirmed that erroneous correspondences are able to be removed by outlier detection using IQR for calculated for each corresponding points. In addition, the proposed method has robustness for the sensor noise to around 20%. The proposal effectiveness is being validated using real images, and applications for obstacle avoidance system or movement control by embedding into in-vehicle camera or mobile robot are investigated.

References [1] R. Hartlry and C. Silpa-Anan, “Reconstruction from two views using approximate calibration, Proceedings of 5th Asian Conference on Computer

116

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Robust H∞ Damping Control of Multi-FACTS Devices for Stability Enhancement of Power Systems with Signal’s Time Delay Fang LIU*&, Min WU*, Yong HE*, Danyun LI*, Ryuichi YOKOYAMA** * Hunan Engineering Laboratory for Advanced Control and Intelligent Automation, School of Information Science and Engineering, Central South University, Changsha 410083, China ** Graduate School of Environmental and Energy Engineering, Waseda University, Tsurumaki- Cho, Shinjuku-ku, Tokyo 162-0041, Japan &Corresponding author: [email protected], Tel: +86-731-88830387 active power, etc., are being used to damp LFO. These strategies are able to damp local oscillation mode, but for the inter-area oscillation mode, they cannot provide effective damping, since the inter-area mode is not observable/controllable directly from the generator’s local signals.

Abstract Abstract: In this paper, a robust H∞ damping controller of multi-FACTS device for a power system is developed with considering the time delay of the remote feedback signals transmitted by wide-area measurement systems (WAMS). In this paper, a free-weighting matrices method based on linear control design approach is presented to design the robust H∞ damping controller to improve the dynamical performance of power systems. Firstly, the linearized reduced-order plant model is established, which efficiently considers the signal’s time delay and the model’s uncertainty. Then, based on the robust control theory, the design of multi-FACTS robust H∞ damping controller is formulated as the standard control problem on delay-dependent state-feedback robust control, which is described by a set of linear matrix inequalities (LMIs) constrains. Finally, the simulation tests are carried out on the 2-area 4-machine power systems. Satisfactory test results verify the correctness of the proposed damping controller and the feasibility of the robust wide-area time-delay control approach. .

In recent years, the wide-area measurement systems (WAMS) are being applied in power systems, which is also one obvious technical feature of the smart grids. It would be wonderful to construct WAMS-based PSS and FACTS supplementary wide-area damping control strategy, which combines the controllable devices and the global monitoring ability of WAMS, to prevent the LFO, especially the inter-area oscillations and thus enhance the global stability of power systems. A method for designing the FACTS damping controller that adds damping on electromechanical oscillations was presented in some literature [3, 5]. However, these methods cannot sufficiently consider the effects of time-varying delay of wide-area signal on the damping performance. Moreover, in the practical large-scale interconnected systems, there is not only one but several wide area damping controllers together to stabilize power oscillations and improve stability of power systems. In such a case, the coordination of multiple controllers should be considered carefully [6-8]. Otherwise, if each controller is designed independently, although each controller can get the effect control performance for the concerned oscillation mode, the interaction among multiple controllers may damage the overall control performance, and the overall stability enhancement maybe not gained.

Keywords: H∞ damping controller, time delay, free-weighting matrices, 2-area 4-machine power systems. 1. INTRODUCTION With the scale-up of electric power networks, the system stability problems become more complex. In particular, the interconnection of power networks and the application of fast excitation systems make the low-frequency oscillations (LFO) become more and more serious, which have been an critical factor that influences the stable and efficient operation of power systems[1-2]. Up to now, power system stabilizer (PSS) and flexible ac transmission systems (FACTS) device, which use local measured signals as feedback signals, such as rotor speed,

For this consideration, a wide area coordinated damping control strategy is proposed in this paper. It utilizes multiple FACTS control devices to construct wide area control network. A robust design approach is correspondingly proposed to simultaneously tune all the considered FACTS wide area damping controllers. The controller interaction is considered sufficiently to reduce or even to eliminate the effect of one controller on another one. Besides these, the varying delay of each wide area

117

Besides this, it is worth to say that in this paper, the designed controller is the typical state-feedback controller, however, in practice, it could be impossible to realize the observation of all the operating states of the large-scale power systems. Therefore, the state observer is introduced to converse such state-feedback control as one kind of output-feedback control. In this paper, the state observer is designed with the classic but practical pole-placement method.

control signal is also considered to maintain control performance.

2. BASIC CONCEPT OF MULTI-FACTS ROBUST H∞ DAMPING CONTROL STRATEGY The basic concept of wide-area robust damping control for multiple FACTS devices can be described as shown in Fig.1. For the large-scale power systems including kinds of FACTS devices (e.g. shunt-type and series-type FACTS devices), the supplementary control function associated to each FACTS devices can be available for the implementation of wide-area damping control to enhance the overall stability of large-scale power systems.

3. ROBUST DESIGN FOR Multi-FACTS DAMPING CONTROLLER 3.1 Problem Formulation When consider the uncertainty and disturbance of power system, the linearized power system with series-type FACTS device (TCSC) and shunt-type FACTS device (SVC) can be described as

Fig.1 Basic framework of robust H∞ damping control of Shunt‐type FACTS‐device

Supplementary control

Series‐type FACTS‐device

POWER SYSTEMS Wide‐Area Signal

Supplementary control

Vmax 1

Vmin

 x (t )  Ax(t )  B1u (t )  B2 (t )  Cx(t )  D2 (t )     z (t )   D1u (t )   

For a given scalar   0 , the performance of the system is defined to be 

K

State Observer O(s)

Thf s 1 Thf s

1 1 Tlf s

(1)

J ( )   z T (t ) z (t )   2 T (t ) (t )dt 0

(2)

Td

The H  robust control problem addressed in this paper

Wide area coordinated damping (WACD) controller

can be stated as described as: for a memory state-feedback controller, find a value for the gain

multi-FACTS devices

K mn , in the control law

From the basic framework as shown in Fig.1, it can be seen that the wide-area control signals should be determined in advance before constructing wide-area damping control. Generally, kinds of operating variables such as line power flow, line current, rotor speed of remote generator, and so on, can be selected as the wide-area control input. The classic residue method can be used to choose the suitable control input [9-10]. In addition, for the wide-area control, the effect of the time delay of the wide area signals on the wide-area control performance should be considered carefully. In this paper, the Pade approximation [11-12] is used to represent the time delay characteristic of the wide area control signals, and the linear robust control theory and design method based on linear matrix inequality is used to handle with the robust control problem of the time delay power system.

u(t )  Kx(t   (t )) Such that for any time-varying delay, satisfying 0   (t )  h   (t )  

(3)

 4

The closed-loop system (6.1) should be asymptotically stable under the condition  (t )  0 ;

J ( )  0 for all non-zero  (t ) under the zero initial condition and a given   0 . So the system(6.1) can be described as  x (t )  Ax(t )  BKx(t   (t ))  B2 w(t )  Cx(t )  D2 w(t )   z (t )      BKx(t   (t ))  

Furthermore, according to the structure of the wide area robust damping controller as shown in Fig.1, it can be seen that the high-pass and the low-pass filters (HPF and LPF) are used to process the wide-area control signals.

(5)

3.2 Bounded Real Lemma (BRL) Before the delay-dependent H∞ robust design for the

118

supplementary wide-area stabilizing control of FACTS devices, a new delay-dependent bounded real lemma (BRL) of the closed-loop system (5) is derived by using FWM approach.

V (xt )  xT (t)Px(t)  xT (t)Px(t) xT (t)Q1x(t)  (1(t))xT (t  (t))Q1 x(t  (t)) t

xT (t)Q2 x(t)  xT (t  h)Q2 x(t  h)  hxT (t)Zx(t)   xT (s)Zx(s)ds t h

 xT (t)(PA  AT P)x(t)  2xT (t)PBKx(t  (t)) 2xT (t)PB2 w(t)  xT (t)(Q1  Q2 )x(t)  (1 )xT (t  (t ))Q1x(t  (t ))  xT (t  h)Q2 x(t  h)

Theorem 1 Consider system (6.1) with u (t )  0 . Given scalar h  0 ,  , and   0 , the system is asymptotically stable and satisfies J ( )  0 for all non-zero  (t ) under the zero initial condition if there exist matrices P  0 , Qi  0, i  1, 2 , Z  0 and X  0 and any appropriately dimensioned matrices N , M such that the following LMIs hold  hT Z  * hR  * *  * *   X 11  * 1    *   *  X 11  * 2    *   *

T2   0  　 　 0 0    I 

1T 0 I * X 13

X 23

t h

 2[xT (t )N1  xT (t  (t)) N2  xT (t  h)N3 ][x(t)  x(t  (t))  

 h T (t) X (t)  

t T t （t ) 2



   A BK

PB2   0  0    2 I 

t h



t

h t 

x (s)Zx(s)dsd

 X 11  * 1    *   *

X 12 X 22

X 13 X 23

* *

X 33 *

 X 11  * 2    *   *

X 12 X 22

X 13 X 23

* *

X 33 *

T

x T ( s ) 

N1  N 2  N3   Z  M1  M 2  M3   Z 

From the inequality of Eq. (10), it can be obtained

(9)

V ( xt )  z T (t ) z (t )   2 T (t ) (t )  1T (t )[  hT R ]1 (t )  

t

t （t )

where P  PT  0 , Qi  QiT  0, i  1, 2 , Z  Z T  0 . Calculating the derivative of

T

 2 (t , s )   xT (t ) xT (t   (t )) xT (t  h)

xT (s)Q1 x(s) T

xT (t  h) 

1 (t )   xT (t ) xT (t   (t )) xT (t  h)  T (t ) 

0 0

0

0 B2 

 (t )   xT (t ) xT (t   (t ))

0 0 D2 

 x (s)Q2 x(s)   T

Q2  M 3  M 3T  hX 33 *

 22  (1   )Q1  N 2  N 2T  M 2  M 2T  hX 22

Proof: Construct the following Lyapunov candidate function:

t

PB2   0  0    2 I 

N 3T  M 1  hX 13  N  M 3T  M 2  hX 23 T 3

12  PBK  N1  N 2T  M 1  hX 12

0 B2 

t  (t )

(10)

11  PA  AT P  N1  N1T  Q1  Q2  hX 11

T 2

t

2T (t, s)22 (t, s)ds   2T (t)(t)

11 12  *  22   * *  *  *

T 1

V (xt )  xT (t)Px(t )  

 (t, s)12 (t, s)ds

where

 22  (1   )Q1  N 2  N  M 2  M  hX 22

 2   0 DK

t  (t )

t h

12  PBK  N1  N 2T  M 1  hX 12

1   C

x(s)ds]

 T (t) X (t)ds

 1T (t)[ hT R]1 (t)  

11  PA  A P  N1  N  Q1  Q2  hX 11

   A BK

t

t （ h t)

(8)

Q2  M 3  M 3T  hX 33 *

T 2

t  (t )

t h

(7)

N 3T  M 1  hX 13  N 3T  M 3T  M 2  hX 23

T

x(s)ds]

 2[xT (t)M1  xT (t  (t))M2  xT (t  h)M3 ][ x(t  (t))  x(t  h)  

where 11 12  *  22   * *  *  *

t

t  (t )

(6)

N1  N 2  0 X 33 N 3   * * Z X 12 X 13 M 1  X 22 X 23 M 2  0 * X 33 M 3   * * Z  X 12

X 22 *

t

 hxT (t)Zx(t)   xT (s)Zx(s)ds



t  ( t )

t h

 2T (t , s)1 2 (t , s )ds

 2T (t , s ) 2 2 (t , s )ds  zT (t ) z (t )

 1T (t )[  hT Z   1T 1  T2  2 ]1 (t )  

V ( xt ) along the solutions

t

t （t )



of system (1) and using the FWMs approach, it obtained:

t  ( t )

t h

119

 2T (t , s ) 2 2 (t , s )ds

 2T (t , s )1 2 (t , s )ds (11)

where 1  C

0 0 D2  ,  2   0 DK

 Y11 Y12 Y13 M1     M2   * Y22 Y23 0  * M3  * Y33   * * LW 1 L   * where      11  AL  LAT  N1  N1T  Q1  Q2  hY11     12  BV  N1  N 2T  M 1  hY12    13  N3T  M 1  hY13        22  (1   )Q1  N 2  N 2T  M 2  M 2T  hY22      23   N 3T  M 3T  M 2  hY23      33  Q2  M 3  M 3T  hY33

0 0

If  i  0,(i  1, 2) and the LMI (6) is true , using Schur complement[14], it is easy to obtain   h Z   0 , T

which means system(1) is asymptotically stable with

 (t )  0 . On the other hand, if  i  0,(i  1, 2) and the LMI (6) is true, it is also easy to obtain , then   hT Z   1T 1   T2  2  0 T 2 T V ( x )  z (t ) z (t )   w (t ) w(t )  0 　 t

That is

z T (t ) z (t )   2 wT (t ) w(t )  V ( xt )

Then the system is asymptotically stable and satisfies

(12)

J ( )  0 for all non-zero

On both sides of Eq. (12), integral from 0 to ∞ with respect to t yields 



0

[ z (t ) z (t )   w (t ) w(t )]dt  V (0)  V () . T

2

T

condition

(13)



0

[ zT (t ) z (t )   2 wT (t ) w(t )]dt  0

and

 (t )

under the zero initial a

u (t )  VL1 x(t )

stabilizing

H  Controller .

It is straightforward to see



(17)

Define

(14)

  diag{P 1

P 1

P 1

1

1

1

  diag{P 3.3 H  Controller Design

P

P

Z 1

I

I

I}

1

P }

Pre- and post-multiply  in (1) by  ;

This section extends Theorem 1 to the design of an H∞ controller for system (1) under control law (3).

pre- and

post-multiply i , i  1, 2 in (2) and (3) by  and make the following changes in the variables:

L  P 1  N i  P 1 N i P 1 , i  1, 2,3  M i  P 1 M i P 1 , i  1, 2,3  Qi  P 1Qi P 1 , i  1, 2

Theorem 2 consider closed-loop system (5). For given scalar h  0 ,  , and   0 , if there exist matrices L  0 ,  Qi  0, i  1, 2 , W  0 and Y  0 and any appropriately

dimensioned matrices

  N , M and V such that the

Y  P 1 XP 1

following matrix inequalities hold:

 11   *  *   *  *   *  *  　　 Y11  * *   *

 12  11 *

 13   23   33

W  Z 1

B2 0

hLAT hV T BT

LC T 0

0

0

0

hB

D2T

* *

 I * *

 hI *

0 I

*

*

*

*

*

Y12

Y13

X 22

X 23

* *

X 33 *

*

*

* *

2

 N1    N2  0  N3   LW 1 L 

T 2

Then, (15)-(17) are derived using the Schur complement. This completes the proof.

0   V D1  0   0   0 15  0   0   I  T

Note that the conditions in Theorem 2 are no longer LMI conditions due to the terms LW 1 L in (16) and (17). As mentioned in [4], we can solve this nonconvex problem by using the idea for solving a cone complementarity problem. So we omit it here.

4. CASE STUDY

(16)

To validate the designed multiple FACTS wide-area damping controller, the four-machine two-area test system [13], which is also the benchmark system for the

120

effectiveness of the design approach for the multi-FACTS damping control.

oscillation damping study, is simulated in this paper. Here, such test system is modified by placing one series-type FACTS device (TCSC) to enhance the interconnected ability and meanwhile one shunt-type FACTS device (SVC) to satisfy the voltage profile of Bus-7, as shown in Fig.2. Such test system is divided into two areas, in which, G1 and G2 belong to Area-1, and G3 and G4 belong to Area-2. The modal analysis indicates that there is a typical inter-area oscillation mode between Area-1 and -2. Such mode is further represented as the oscillation between G1 (in Area-1) and G3 (in Area-2). Furthermore, the results of the residue analysis indicate that if the current in line 9-10 is selected as the wide-area control input for the multiple FACTS wide-area damping controller, the high controllability can be achieved. Therefore, the current in line 9-10 is chosen, and the formed FACTS-WARD controller has been simply represented in Fig 6.1. The controller parameter can be obtained according to the robust design method presented in the above section.

(a)

(b) Fig.2 Modified four-machine two-area test system installed with multi-FACTS devices

Fig.3 Dynamic response of generator rotor speed of the four-machine two-area test system, (a) without multi-FACTS damping controller; (b) with multi-FACTS damping controller

To reveal the damping performance of the designed multi-FACTS damping controller, a big disturbance (three-phase-to-ground fault nearby bus 13 of line 13-9) is carried out at 0.1 s with the duration of 0.1 s. Fig.3 shows the dynamic responses of the rotor speed of each generator located in the different areas of the test system. From Fig.3(a), it can be clearly seen for the test system without the multi-FACTS damping controller, the disturbance exists the typical inter-area oscillation between Area 1 (G1 and G2) and Area 2 (G3 and G4). Different from this, for the test system installed with the designed damping controller, from Fig.3(b), it can be clearly seen that such oscillation is damped quickly, which validates the

Furthermore, Fig.4 represents the dynamic response of the relative angle between G1 (located in Area 1) and G3 (located in Area 2). From this, it further indicates the obvious feature of the inter-area oscillation in the interconnected system. But when the test system is installed with multi-FACTS damping controller, such serious inter-area oscillation mode is damped effectively. In addition, Fig.5 further represents the dynamic response of the power flow in the tie transmission line. From this, it can be also seen that the serious power oscillation in the tie-line is damped very well with the application of

121

power systems, it may include multiple FACTS devices. Here, these FACTS devices are further developed to implement wide area damping control strategy together. The basic concept and the controller structure are discussed briefly, and then a robust design approach is presented to handle with multiple controllers design and the multiple time delays of wide area control signals. Two case studies validate the presented control concept and controller design approach, and further indicates the good damping performance under various operating conditions and when suffering various time-varying delays of different wide area signals.

multi-FACTS wide area damping control strategy.

70 Without WACD controller With WACD controller Relative angle between G1 and G3/ o

60

50

40

30

20

10

0

0

2

4

6

8

10 t/s

12

14

16

18

20

ACKNOWLEDGEMENT

Fig.4 Dynamic response of relative angle between G1 and G3 located in

This work was supported in part by the National Natural Science Foundation of China (No.61304092), in part by the Research Fund for the Doctoral Program of Higher Education of China(No.20130162120022) and in part by the Hunan Provincial Natural Science Foundation of China (No.13JJ6004).

the four-machine two-area test

16

Power flow in line 3-20/pu

15

14

References [1] P.A. Park. A delay-dependent stability criterion for systems with uncertain time-invariant delays, IEEE Trans. Automatic Control, vol.44, pp. 876-877, 1999. [2] Z. Y. Yuan, X. Tao, Y. C. Zhang, C. Lang, P. N. Markham, R. M. Gardner, and Y. L. Liu. Inter-area oscillation analysis using wide area voltage angle measurements from FNET, in IEEE Power & Energy Society Gen-eral Meeting, 2010. [3] A. Bose. Smart transmission grid applications and their supporting infrastructure, IEEE Trans. Smart Grid, vol. 1, no. 1, pp. 11-19, June 2010. [4] F. Liu, M. Wu, Y. He, and R. Yokoyama. Delay-dependent Robust Stability Analysis for Interval Neural Networks with Time-varying Delay. IEEJ Transactions on Electrical and Electronic Engineering, ISSN: 1931-4973, 2011.07, 6(4), 345-352. [5] L. El Ghaoui, F. Oustry, and M. AitRami. A cone complementarity linearization algorithm for static output-feedback and related problems, IEEE Trans. Automatic Control, vol. 42, pp. 1171-1176, 1997 [6] I. Kamwa, R. Grondin, and Y. Hebert. Wide-area measurement based stabilizing control of large power systems-a decentralized/hierarchical approach. IEEE Transaction on Power Systems. Vol. 16, No. 1, pp. 136–153, Feb. 2001. [7] E. Johansson, K. Uhlen, A.B. Leirbukt, P. Korba, J.O. Gjerde, and L.K. Vormedal. Coordinating power

13

12

11 Without WACD controller With WACD controller 10

0

2

4

6

8

10 t/s

12

14

16

18

20

Fig.5 Dynamic response of power flow in the tie-line 3-20 of the four-machine two-area test system with or without multi-FACTS damping controller

To evaluate the effects of signal time delay on the wide area damping performance, different time delays are acted on both the SVC and the TCSC wide area signal. Here, the delays of both the control-input and the control-output signals are wholly equivalent to such defined time delays. Fig.6 shows the dynamic responses of the test system with different time delays of the wide area signals. From this, it is clearly seen that even there exists various delays for multi-FACTS wide area control signals, the designed damping controller can maintain the good damping performance on the inter-area oscillation.

5. CONCLUSION In this paper, a wide area coordinated robust damping control strategy is proposed to improve the overall stability of large-scale power systems. For the large-scale

122

[11] N.R. Chaudhuri, B. Chaudhuri, S. Ray, and R. Majumder. Wide-area phasor power oscillation damping controller: a new approach to handling time-varying signal latency. IET Generation, Transmission & Distribution. Vol. 4, No. 5, pp. 620-630, May 2010. [12] Y. Yuan, Y.Z. Sun, and L. Cheng. Design of delayed-input wide-area FACTS controller using genetic algorithm. IEEE Power Engineering Society General Meeting, 2007. [13] P. Kundur. Power Stability and Control. New York: McGraw-Hill, 1994. [14] Gao H, Lam J, Wang C, Wang Y. Delay-dependent output-feedback stabilization of discrete-time systems with time-varying state delay. IEE Proceedings Control Theory and Applications 2004; 151: 691–698.

oscillation damping control using wide area measurements. IEEE/PES Power Systems Conference and Exposition, 2009. [8] J.M. Davalos Ramirez, V. R.J. Valenzuela. Coordination of FACTS-based stabilizers for damping oscillations. IEEE Power Engineering Review. Vol. 20, No. 12, pp. 46-49, Dec. 2000. [9] A. Heniche and I. Kamwa. Assessment of two methods to select wide-area signals for power system damping control. IEEE Transactions on Power Systems. Vol. 23, No. 2, pp. 572-581, May 2008. [10] H.F. Wang, F.J. Swift, and M. Li. Indices for selecting the best location of PSSs or FACTS-based stabilisers in multimachine power systems: a comparative study. IEE Proceedings of Generation, Transmission and Distribution. Vol. 144, No. 2, pp. 155-159, Mar. 1997.

123

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Novel Realization of Adaptive Sparse Sensing using Reweighted Zero-Attracting Least Mean Forth Algorithm Guan GUI, Shin-ya MATSUSHITA and Li XU Department of Electronics and Information Systems, Akita Prefectural University, Akita, 015-0055 Japan have been proposed to find the suboptimal sparse solution. It is well known that the CS provides a robust framework that can reduce the number of measurements required to estimate a sparse signal. Many NSS algorithms and their variants have been proposed to deal with CS problems. They mainly fall into two basic categories: convex relaxation (basis pursuit de-noise, BPDN [6]) and greedy pursuit (orthogonal matching pursuit, OMP [7]). Above NSS based CS methods are either high complexity or low performance, especially in the case of low signal-to-noise (SNR) regime. Indeed, it was very hard to adaptive tradeoff between high complexity and good performance.

Abstract Nonlinear sparse sensing (NSS) techniques have been adopted for realizing compressive sensing in many applications such as Radar imaging. Unlike the NSS, in this paper, we propose an adaptive sparse sensing (ASS) approach using reweighted zero-attracting normalized least mean fourth (RZA-NLMF) algorithm which depends on several given parameters, i.e., reweighted factor, regularization parameter and initial step-size. First, based on the independent assumption, Cramer Rao lower bound (CRLB) is derived as for the performance comparisons. In addition, reweighted factor selection method is proposed for achieving robust estimation performance. Finally, to verify the algorithm, Monte Carlo based computer simulations are given to show that the ASS achieves much better mean square error (MSE) performance than the NSS.

0.7 0.6 0.5 Magnitude

Keywords: Reweighted Zero-Attracting Normalized Least Mean Fourth (RZA-NLMF), Compressive Sensing (CS), Sparse Constraint, Adaptive Sparse Sensing (ASS), Nonlinear Sparse Sensing (NSS).

0.4 0.3 0.2

1. INTRODUCTION 0.1

Compressive sensing [1], [2] has been attracting high attentions in compressive Radar/sonar sensing [3], [4] due to many applications such as civilian, military, and biomedical. The main task of CS problems can be divided into three aspects as follows: 1) sparse signal learning: The basic model suggests that natural signals can be compactly expressed, or efficiently approximated, as a linear combination of prespecified atom signals, where the linear coefficients are sparse as shown in Fig. 1 (i.e., most of them zero); 2) random measurement matrix design. It is important to make a sensing matrix which allows recovery of as many entries of unknown signal as possible by using as few measurements as possible. Hence, sensing matrix should satisfy the conditions of incoherence and restricted isometry property (RIP) [5]. Fortunately, some special matrices (e.g., Gaussian matrix and Fourier matrix) have been reported that they are satisfying RIP in high probably; 3) sparse reconstruction algorithms. Based on previous two steps, many sparse reconstruction algorithms

0 0

5

10

15 Taps index

20

25

30

Fig. 1 A typical example of sparse structure signal.

In this paper, we propose an adaptive sparse sensing (ASS) method using reweighted zero-attracting normalized mean fourth error algorithm (RZA-NLMF) [8] to solve the CS problems. Different from NSS methods, each observation and corresponding sensing signal vector will be implemented by the RZA-NLMF algorithm to reconstruct the sparse signal during the process of adaptive filtering. According to the concrete requirements, complexity of the proposed ASS method could be adaptive reduced but without sacrificing much recovery performance. The effectiveness of our proposed method is confirmed via computer simulation when comparing with NSS.

124

The remainder of the paper is organized as follows. Basic CS problem is introduced and typical NSS method is presented in Section 2. In section 3, ASS using RZA-NLMF algorithm is proposed for solving CS problems and its derivation process is highlighted. Computer simulations are given in Section 4 in order to evaluate and compare performances of the proposed ASS method. Finally, our contributions are summarized in Section 5.

reconstructed correctly by NSS methods, e.g., BPDN [6] and OMP [7]. Take the BPDN as for the example to illustrate NSS realization approach. Since the sensing with positive matrix X satisfies RIP of order parameter  K  (0, 1) , i.e., X  RIP(K ,  K ) if (1   K ) h

2 2

 Xh

2 2

2

 (1   K ) h 2 ,

(4)

holds for all h having no more than nonzero coefficients. Then the unknown sparse vector h can be reconstructed by BPDN as 2 1  hnss  arg lim  y  Xh 2   h 1  , h 2 

(5)

where  denotes a regularization parameter which balances the mean-square error (MSE) term and sparsity of h . If the mutual interference of sensing matrix X can be completely removed, then the theoretical Cramer-Rao lower bound (CRLB) of the NSS can be derived as [9][10]



CRLB{hnss }  E hnss  h Fig.2 RZA-LMSF algorithm for realizing ASS.

finite-length discrete signal vector s  [s 1 , s 2 ,  , s N ]T can be sparse represented in a signal domain D , that is

ym  hT xm  zm ,

N

(1)

where h  [h1 , h 2 ,  , h N ]T is the unknown -sparse coefficients vector ( K N ), and D is an N  N orthogonal basis matrix with {di , i  1, 2, , N } as its columns. Take a random measurement signal matrix W and then the received signal vector y  [y 1 ,  , y m ,  , y M ]T can be written as (2)

where X WD denotes a M  N random sensing matrix as  xT1   x 11         X  x Tm    x m 1         xT  x  M   M1

 x 1n  x 1N        x mn  x mN  ,        x Mn  x MN 

(6)

(7)

for m  1, 2,  , M . The objective of ASS is to adaptively estimate the unknown sparse vector h using the sensing signal vector x m and the observed signal y m . Different from NSS approaches, we proposed an alternative ASS method using RZA-NLMF algorithm as shown in Fig. 2. Assume the ym (n )  x Tm h (n ) is an estimated observed signal which depends on signal estimator h (n ) and hence the -th observed signal error as em (n )  ym  ym (n ) . Notice that the em (n) is in correspondence with the -th iterative error when using -th sensing signal vector x m and m  mod(n , M ) . Notice that the mod() denotes a modulo function, for example, mod(10, 3)  1 and mod(11, 3)  2 . First of all, the cost function of RZA-NLMF algorithm is constructed as 1 N G (n )  em4 (n )  ass  i 1 log 1   hi , (8) 4 where ass  0 is a regularization parameter which trades off the sensing error and coefficients vector sparsity.   0 denotes a reweighted factor which enhances to exploit the signal sparsity at each iteration. A figure example to show the relationship between reweighted factors and sparse constraint strength is given in Fig. 3. According to the cost function (8), the

i 1

y Ws  z WDh  z  Xh  z

.

We reconsider the above system model (2) with respect to adaptive sensing case. At observation side, -th observed signal y m can be written as

a

s   di hi  Dh ,

2 n

3. ADAPATIVE SPARSE SENSING

2. NONLINEAR SPARSE SENSING Assume

2

  KN

(3)

and z  [z 1 ,  , z m ,  , z M ]T is an additive white Gaussian noise (AWGN) with distribution  (0,  n2I M ) and I M denotes an M  M identity matrix. From the perspective of CS, the sensing matrix X satisfies the restricted isometry property (RIP) in overwhelming probability [5] so that the sparse signal h can be

125

corresponding update equation can be derived as G (n ) h (n  1)  h (n )  iss h (n )

0.9

issem3 (n )x m xm

2 2

x

2 m 2

 e (n ) 2 m





 (n )em (n )x m  sgn  h (n )  ass  2 1 xm 2

 sgn h (n )  1   h (n )

Strength of sparse penalty

 h (n ) 

1

(9)

h (n )  , h (n )

where   iss   is a parameter which depends on initial step-size iss , regularization parameter  and threshold  , respectively. In the second term of (9), if coefficient magnitudes of h (n ) are smaller than 1  , then these small coefficients will be replaced by zeros in high probability [11]. Here, it is worth noting that ass (n ) is a variable step-size

issem2 (n )

ass (n ) 

2

 em2 (n ) 2

xm

xm

2

em2 (n )  1





5iss

2 2 4 n

,



9iss n2 2  2 2



0.6 0.5 0.4 0.3 0.2

0 -1

weak penalty -0.5

0

0.5

1

Value of coefficient Fig. 3 Sparse constraint strength comparison using different reweights.

4. COMPUTER SIMULATIONS In this section, the proposed ASS approach using RZA-NLMF algorithm is evaluated. For achieving average performance, 1000 independent Monte-Carlo runs are adopted. For easy evaluating the effectiveness of the proposed approach, signal representation domain D is assumed as an identity matrix I N N and unknown signal is set as sparse directly. Sensing matrix is equivalent to random measurement matrix, i.e., X W . For ensuring X satisfies the RIP, W is set as random Gaussian matrix [5]. Then, sparse coefficient vector h equals to s . The detail simulation parameters are listed in Tab. 1. Notice that each nonzero coefficient of h follows random Gaussian distribution as  (0,  2 ) and their positions are randomly allocated within the signal length of h which is subject to E {|| h ||22 }  1 , where ∙ denotes the expectation operator. The output signal-to-noise ratio (SNR) is defined as 20log ⁄ , where 1 is the unit transmission power. All of the step sizes and regularization parameters are listed in Tab. I. The estimation performance is evaluated by average mean square error (MSE) which is defined by

(11)

which is a variable step-size (VSS) which is adaptive change as square sensing error em2 (n ) , smaller error incurs the smaller step-size to ensure the stability of the gradient descend while larger error yields larger step-size to accelerate the convergence speed of this algorithm [12]. According to the update equation in (9), our proposed ASS method can be concluded in Algorithm 1, where   0 is a given error tolerance and nmax is a given maximum iteration number. The CRLB of the proposed ASS can be obtained as CRLB{hass }=E hass  h

0.7

(10)

,

iss 2

0.8

0.1

which depends on three factors: initial step-size  iss , input signal x m and update iterative error em (n) . Since  iss is given initial steps-size and x m is random scaling input signal, hence, ass in Eq. (10) can also be rewritten as

ass (n ) 

=0 =2 =20 =200 =2000

stronge penalty

 2NK . 2  n4  6 iss n2 27 iss

(12)

About detailed derivation of the CRLB, interested authors are suggested to refer [13].







Average MSE h (n ) : E h  h (n )

Input: X , y ,  iss ,  and  Output: .

2 2

,

(13)

where h and h (n ) are the actual channel vector and its -th iterative adaptive channel estimator, respectively. According to our previous work [8], regularization parameter for RZA-NLMF is set as   5 108 so that it can exploit signal sparsity robustly. Since the RZA-NLMF-based ASS method depends highly on the reweighted factor  , hence, we first select the reasonable factor  by virtual of Monte Carlo. Later, we compare the proposed method with two typical NSS ones, i.e., BPDN [6] and OMP [7].

Initialize: h (0 )  0 ;

n 1 ; or n  n max Do While h (n  1)  h (n )   2 (1) Determine xm and y m with m  mod(n , M )  1 ; (2) Calculate error em (n ) as e m (n )  y m  x Tm h (n ) ; (3) Update h (n 1) with update Eq. (9); (4) n  n  1 End Algorithm 1. Realization of ASS using RZA-NLMF.

126

It is well known that ℓ -norm normalized least mean fourth (L0-NLMF) for CS can achieve optimal solution but it is a NP hard problem in practical applications such as noise environment [2]. One can find that RZA-NLMF reduces to L0-NLMF when the reweighted factor approaches to infinity. Due to the noise interference, we should select the suitable reweighted factor which can not only exploit signal sparsity but also can mitigate noise interference effectively. Hence, reweighted factor of the RZA-NLMF is selected empirically. By means of Monte Carlo method, performance curves of the proposed ASS method with different reweighted factors ε ∈ 2,20,200,2000,20000 with respect to different number of nonzero coefficients ∈ 2,6,10 and different SNR regimes (5dB and 10dB) are depicted in Figs. 4~7.

Table 1. Simulation parameters. Parameters

Values

Signal length

40

Measurement length

20

Sensing matrix

Random Gaussian distribution ∈ 2,6,10

No. of nonzero coefficients Distribution of nonzero

Random Gaussian

coefficients Signal-to-noise ratio (SNR)

(0dB, 12dB)

Initial step-size:

1.5 5

Regularization parameter: Re-weighted factor:

10

2000

4.1. Reweighted factor selection Since the RZA-NLMF algorithm depends highly on reweighted factor. Hence, selection of the robust reweighted factor for different noise environments and different signal sparsities is typical important step for the RZA-NLMF algorithm. RZA-NLMF ( =2) RZA-NLMF ( =20) RZA-NLMF ( =200) RZA-NLMF ( =2000) RZA-NLMF ( =20000)

MSE

MSE

-1

10

Signal lenght: N=40 No. nonzero: K=2 SNR=5dB

-1

10

RZA-NLMF ( =2) RZA-NLMF ( =20) RZA-NLMF ( =200) RZA-NLMF ( =2000) RZA-NLMF ( =20000)

Signal lenght: N=40 No. nonzero: K=2 SNR=10dB

-2

10

-3

10 -2

10

-4

10

0

1000

2000

3000 Iterations

4000

5000

6000

0

0.5

1

1.5

2

2.5

3 4

Iterations

x 10

Fig. 4 RZA-NLMF performance verses reweighted factors ( =2 and

Fig. 6 RZA-NLMF performance verses reweighted factors ( =6 and

SNR=5dB).

SNR=10dB). RZA-NLMF ( =2) RZA-NLMF ( =20) RZA-NLMF ( =200) RZA-NLMF ( =2000) RZA-NLMF ( =20000)

-1

-1

10

Signal lenght: N=40 No. nonzero: K=2 SNR=10dB

-2

10

MSE

MSE

10

RZA-NLMF ( =2) RZA-NLMF ( =20) RZA-NLMF ( =200) RZA-NLMF ( =2000) RZA-NLMF ( =20000)

Signal lenght: N=40 No. nonzero: K=12 SNR=10dB

-2

10

-3

10

-3

10

-4

10

-4

10

0

0.5

1

1.5

Iterations

2

2.5

3

0

4

x 10

0.5

1

1.5

Iterations

2

2.5

3 4

x 10

Fig. 5 RZA-NLMF performance verses reweighted factors ( =2 and

Fig. 7 RZA-NLMF performance verses reweighted factors ( =12 and

SNR=10dB).

SNR=10dB).

127

The existing big performance gap between ASS and NSS is that ASS using RZA-NLMF not only exploits the signal sparsity but also mitigates the noise interference using high-order error statistics for adaptive error updating. On the other hand, we can also find that ASS depends on the signal sparseness. That is to say, for sparser signal, ASS can exploit more signal structure information as for prior information and vice versa. In the second experiment, number of nonzero coefficients is fixed as K 2 as shown in Fig. 9. It is easy to find that our proposed ASS is much better than conventional NSS as the SNR increasing.

Under the simulation setup considered, RZA-NLMF using ε 200 0 can achieve robust performance in different cases as shown in Figs. 4~7. From the four figures, one can find that sparser signal requires larger reweighted factor but no more than 20000 in this system. This is concise with the fact that stronger sparse penalty not only exploits more sparse information but also mitigates more noise interference. 4.2. Performance comparisons with NSS Two experiments of ASS are verified in performance comparisons with conventional NSS methods (e.g., BPDN [6] and OMP [7]). In the first experiment, ASS method is evaluated in the case of SN R  10 dB as shown in Fig. 8. On the one hand, according to this figure, we can find that the proposed ASS method using RZA-NLMF algorithm achieves much lower MSE performance than NSS methods and even if its CRLB.

5. CONCLUDING REMARKS In this paper, we proposed an ASS method using RZA-NLMF algorithm for dealing with the CS problems. First, we decided the reweighted factor and regularization parameter for the proposed algorithm by virtual of Monte Carlo method. Later, based on update equation of the RZA-NLMF, CRLB of ASS was also derived based on the random independent assumptions. Finally, several representative simulations have been given to show that proposed method achieves much better MSE performance than NSS with respect to different signal sparsity, especially in the case of low SNR regime.

0

10

-1

Average MSE

10

RZA-NLMS (ASS)

-2

10

BPDN (NSS)

References

OMP (NSS)

[1] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, 2006. [2] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289–1306, Apr. 2006. [3] R. Baraniuk, “Compressive radar imaging,” in IEEE Radar Conferencn, MA, USA, 17- 20 April 2007, 2007, pp. 128–133. [4] M. Herman and T. Strohmer, “Compressed sensing radar,” in IEEE Radar Conference, Rome, Italy, 2-5 Sept. 2008, 2008, no. 2, pp. 1–6. [5] E. J. Candes, “The restricted isometry property and its implications for compressed sensing,” Comptes Rendus Math., vol. 1, no. 346, pp. 589–592, May 2008. [6] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition by basis pursuit,” SIAM J. Sci. Comput., vol. 20, no. 1, pp. 33–61, 1998. [7] J. A. Tropp and A. C. Gilbert, “Signal recovery from random measurements via orthogonal matching pursuit,” IEEE Trans. Inf. Theory, vol. 53, no. 12, pp. 4655–4666, 2007. [8] G. Gui, A. Mehbodniya, and F. Adachi, “Adaptive sparse channel estimation using re-weighted

CRLB (NSS)

-3

10

-4

10

-5

10

2

3

4

5

6

7

8

9

10

K Fig. 8 Performance comparisons verses signal sparisty.

0

10

-1

Average MSE

10

-2

10

-3

10

-4

10

BPDN (NSS) OMP (NSS)

-5

10

RZA-NLMS (ASS)

0

CRLB (NSS) 2

4

6

8

10

12

SNR (dB) Fig. 9 Performance comparisons verses SNR.

128

[12] G. Gui, L. Dai, S. Kumagai, and F. Adachi, “Variable earns profit : Improved adaptive channel estimation using sparse VSS-NLMS algorithms,” in in IEEE International Conference on Communications (ICC), Sydney, Australia, 10-14 June 2014, 2014, pp. 1–5. [13] G. Gui, L. Xu, and F. Adachi, “RZA-NLMF algorithm based adaptive sparse sensing for realizing compressive sensing problems,” EURASIP J. Adv. Signal Process., 2014.

zero-attracting normalized least mean fourth,” in 2013 2nd IEEE/CIC International Conference on Communications in China (ICCC), Aug.12-14, 2013, Xian, China, 2013. [9] L. Dai, Wang ， Zhaocheng, and Z. Yang, “Compressive sensing based time domain synchronous OFDM transmission for vehicular communications,” IEEE J. Sel. Areas Commun., vol. 31, no. 9, pp. 460–469, 2013. [10] L. Dai, Z. Wang, and Z. Yang, “Spectrally efficient time-frequency training OFDM for mobile large-scale MIMO systems,” IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 251–263, 2013. [11] Y. Chen, Y. Gu, and A. O. Hero III, “Sparse LMS for system identification,” in IEEE International Conference on Acoustics, Speech and Signal Processing, April 19-24, 2009,Taipei, Taiwan, 2009, no. 3, pp. 3125–3128.

129

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Practice of Active Learning Based on Craftsmanship Liang Xiao＊ (Tokyo Univ. of Tech.), Zhejun Fang (Tokyo Univ. of Tech.), Jie Zhang (Waseda Univ.), Jinhua She (Tokyo Univ. of Tech.), Yasuhiro Ohyama (Tokyo Univ. of Tech.)

* Graduate School of Bionics, Computer and Media Sciences, Tokyo University of Technology Tokyo, 192-0982, Japan ** School of Computer Science, Tokyo University of Technology Tokyo, 192-0982, Japan During the advancement of university education reform, many universities introduce a teaching method called active learning. It is not a way that the teachers teach and the students passively learn, but a way that incorporates research, discussion, and presentation, etc., It concentrates more on students’ active participation. This paper presents an active-learning practice which is based on constructing a model airplane. One purpose of this airplane-making course is to let students understand the basic of manufacturing, be interest in mechatronics, and be more self-motivated in the rest of 2 years’ study. Another purpose is to improve students’ teamwork and communication ability by making a model airplane cooperatively. Keywords: Engineering based education, active learning, control engineering be interest in mechatronics, and be more self-motivated

1. Introduction Conventional

in the rest of 2 years’ study. Another purpose is to

education

introduces

report

improve students’ teamwork and communication ability

and

by cooperative practical manufacturing.

discussion to promote students’ desire of learning and improve the degree of understanding.

In recent years,

2. Construction of a Radio-controlled (RC) Model Airplane

research on a new education style, which is called active learning, is becoming popular in college education around the world. And this marks the beginning of

In Tokyo University of Technology, the grade-2

research on student-centered education. Active learning

students in the computer science department are

takes not “what the teachers should teach?”, but “what

assigned seven courses. For students who belong to

the students can learn?” as its principle of education.

mechatronics department, a course which is called

Therefore, the class is not limited to lectures, but purses

“mechatronics lab (laboratory)” is provided as an entry

a style that students can actively participate.

course of the mechatronics control theory. This course

This study focuses on mechatronics course of grade-2 students

in

Tokyo

University

of

Technology.

includes four themes, and this study focuses on one of

By

them: construction of a model airplane. This theme is

constructing a model airplane, one purpose of this course

divided into three lessons, each includes 2 periods (1.5

is to let students understand the basic of manufacturing,

hours *2).

Students are divided into 2 groups, each

Figure 2 List of learning contents

Figure 1. Conventional education.

130

Figure 4 The basic structure of an airplane.

An airplane includes three main parts: the body, the main wings, and the tail part (Figure 4).

Figure 3 Examples of aircrafts.

All forces that act on an airplane are shown in Figure 5. The engine produces thrust that pushes the airplane

group has 3. The first lesson introduces basic airplane

forward, and therefore generates high speed airflow on

structure, air dynamics, and flight principles. Based on understanding,

a

computation

and

wings. The difference of air pressure on airfoil surface

simulation

produces left. When the lift is greater than the weight of

application Scilab, which is free, is used to explore the

the airplane, the airplane rises. During a horizontal

relationship between lift and the airplane speed/airfoil

flight, the lift equals the weight of the airplane.

surface area; The second lesson is the construction the model airplane based on previous introduction and

4. Understanding and Verification Using Scilab

pre-prepared materials and tools; The last one is

For further understanding of the flight principle,

completing the airplane, computer flight simulation and

Scilab is used to explore the relationship among lift, the

on-site flight test.

speed, and the area of the airfoil. The simulation block diagram is shown in Figure 6, and the curve shown in

3. Airplane Basics and Flight Principles

Figure 7 shows the relationship between the airplane

An airplane, examples are shown in Figure 3, is a kind

speed and generated lift. By encouraging students

of machine that flies in the air. It can be classfied into

plotting other curves, such as lift vs the airfoil area with

two main types: LTA (lighter-than-air aircraft for

constant speed, airfoil area vs speed with constant lift,

example, ballon and airships ), and HTA (heavier- than

they will have a deeper understanding of the flight

-air aircraft, for example, airplane and helicopter). Base

principle.

on its wing type, they can be divided into two classes: fix

The lift

wing airplane (airplane in general sense, for example, the

Boeings

),

and

rotorcraft

(for

example,

the

L

holicopters).

Figure 5 Illustration of forces act on an airplane.



1 V 2 SCL ·················································· (1) 2

Figure 6 Simulation block diagram in Xcos of Scilab.

131

Figure 8 The design of the model airplane.

such work for most of those students, the construction Figure 7 Simulation result: speed (m/s) vs lift (N).

cost a long time. However, they all finished their work as scheduled.

where CL： Coefficient of Lift ρ ：Density（sea level air density: 1.2250

6. Flight Test kg/m3）

The flight test was conducted on the playground of

V : Relative speed

Hichioji campus, Tokyo University of Technology. Each

S ：Surface area

group tested their model plane one by one.

L ：Lift

7. Conclusion By now, this course has opened two times, a total of 4

5. Construction of the Model Airplane

groups registered. All students seem intersted in it and

The overall design of the model airplane is shown in

actively pacitipate in each step.

Figure 8. The body and wings are made of Styrofoam

On ther other hand,

when the model plane flew during the flight test, they

board for its light-weight and comparably good strength.

experienced the excitement they never had before. This

The wings span is 90 cm, the body length is 60 cm, and

arouses their interest and also curiocities. A questionare

the total weight is about 500 grams. Figure 9 shows the

is planed to investiage whether this course has any

scene during the class: there are two groups, each group

positive effect on the specialize courses followed by.

includes three students. All students showed a strong interest in this course. Because it is the first time to do

References

Figure 9 Scene of construction of the airplane in the

Figure 10 Flight test.

laboratory.

132

(１) Wikipedia ： Airplanes ， [online

available]

http://ja.wikipedia.org/

wiki/%E8%88%AA%E7%A9%BA%E6%A9%9F (access：2014.4.14） (２) Katsuhiko Aoki, Practice of organized team-activities on engineering foundations education, Journal of JSEE, pp 23-28, 2009 (In Japanese) (３) Tetsuya Taniguchi, Evaluation of problem-solving type active learning, Journal of JSEE，61－3(2013)，10－13，2013 (In Japanese)

133

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

The energy and transmission optimization based on dynamic spectrum sensing in cognitive radio sensor networks Yi LI*, Jun PENG*, Xiaoyong ZHANG*1, Kaiyang LIU*, Fu JIANG* * School of Information Science and Engineering, Central South University Changsha, Hunan, China 400075 [1e-mail: [email protected]]

technology has been proposed to solve this problem for improving the reliability of the spectrum sensing. In cooperative spectrum sensing, each node makes a local decision and reports it to a fusion center for a final decision according to some fusion rules[5-6].

Abstract As the inherent energy constraint problem, the energy efficiency optimization has become a critical issue for cognitive radio sensor networks. To address this issue, a joint energy efficiency optimization method for spectrum sensing and data transmission is investigated in this paper. Then a dynamic censored based cooperative spectrum sensing scheme is employed to decide when to stop sensing. This scheme could shorten sensing time and save unnecessary spectrum sensing energy. Aiming at jointly optimizing the energy efficiency of spectrum sensing and data transmission, the distortion constrained probabilistic transmission scheme is utilized. Numerical simulations compare the performance of the proposed scheme with a fixed spectrum sensing scheme in different scenarios. It is shown that significant energy efficiency improvement could be achieved in this paper.

Since the inherent energy constraint character of cognitive sensor nodes, the energy efficiency becomes a critical challenge of CRSN to prolong the lifetime of networks. Energy efficient spectrum sensing approach can be employed to relieve the shortage of energy. For instance, maleki et al. propose a combined sleeping and censoring scheme in [7], a censored truncated sequential sensing scheme in [8] and discuss the optimal hard strategy in fusion center [9] to reduce the whole network’ or each node’ energy consumption in spectrum sensing. Najimi et al. [10] bring forward a joint sensing nodes and decision nodes selection method to minimize the total energy consumption in spectrum sensing subject to a global detection performance constraint. Hao et al. [11] formulated the multi-channel spectrum sensing problem as a coalition formation game, and the stable partition of this coalition game can be achieved by a proposed distributed algorithm. Zahmati et al. [12] adopt a hybrid sensing algorithm to find the optimal sensing period based on the property of networks. This scheme avoids unnecessary sensing tasks and saves the spectrum sensing energy.

Keywords: Cooperative Spectrum Sensing, Energy Efficiency, Dynamic Censoring, Cognitive Radio Sensor Networks.

1. INTRODUCTION With the rapid development of wireless technologies, the limited frequency spectrum available for wireless sensor network applications has been heavily crowded[1]. Therefore, the concept of cognitive radio sensor networks (CRSN) has been proposed to address this problem. Equipped with cognitive radio technologies, sensor nodes can dynamically using the idle resources in authorized spectrum[2]. Benefiting from that, cognitive radio sensor networks have been widely applied in many fields, such as telemedicine, home-security monitoring, emergency response networks[3].

However, the above researches all focus on reducing energy consumed in spectrum sensing period. In fact, energy efficiency can be achieved through effectively data transmission as well. The scholars have conducted some researches to this problem. Jain et al. [13-14] have investigated the energy efficiency of lossy transmission of correlated CEO and bivariate Gaussian source in wireless sensor networks. And an energy lower bound was derived using cut-set arguments before comparison to analog transmission and separate source-channel coding. Zhang et al. [15] investigated the energy efficiency in joint source and channel sensing. Within bounded distortion, the total energy efficiency could be optimized. Naeem at

In cognitive radio sensor networks, spectrum sensing is an important technique to determine the white spectrum[4]. However, the result of spectrum sensing may be affected by many factors such as multipath fading and shadowing. Therefore, cooperative spectrum sensing

134

scheme is used to determine the vacant channels for opportunistic spectrum access. The main objective of spectrum sensing is to realize effective and efficient exploration of spectrum resources.

el. [16] modeled the power allocation of spectrum and source sensing as a nonlinear programming problem, and used Charnes-Cooper transformation technique to convert the problem into a convex optimization problem. However, these papers don’t take the cooperative or energy efficiency spectrum sensing approach into consideration, and are easily affected by wireless fading and shadow effect. Hence an important issue arises: how to jointly optimize the energy efficiency of both spectrum sensing and data transmission of cognitive radio sensor networks? To address this issue, a joint energy efficiency optimization of both spectrum sensing and data transmission is investigated in this paper. The major contributions of this paper can be summarized as follows:  A dynamic censored based cooperative spectrum sensing approach is proposed to decide when to stop sensing. This spectrum sensing approach could shorten sensing time and to save the unnecessary spectrum sampling energy constraints on the detection and transmission performance.

Fig.1 An overview of considered system model.

The Fig.1 depicts the subtle interplay between spectrum sensing and application source sensing under interference, distortion and power constraint. The probability of effective transmission is the used to bond the energy efficiency problem in two tasks.

 To jointly optimize the energy efficiency of spectrum sensing and data transmission, the effective transmission probability is used to bond the energy efficiency problem in two periods. And the data transmission is constrained by distortion as well.

A cognitive radio sensor network with K nodes is considered, where CRSN nodes gain spectrum access if the primary user is not active within a certain band. Each CRSN node employs periodic time slots of length T for spectrum sensing and data transmission. The spectrum sensing and transmission scheme in this paper is presented as Fig.2. It is assumed that the spectrum sensing is always performed ahead of the data transmission. Denote tsj to be the time of spectrum sensing, and tdj to be the time of data transmission of the j-th CRSN node, where j ∈[1, K ].

 The performance evaluation and comparison can reveal the effectiveness and superiority of the proposed joint energy efficiency optimization proposed in this paper. The rest of this paper is organized as follows. In Section 2, the system model as well as the dynamic censored based cooperative spectrum sensing approach has been described. In Section 3, mathematical expressions for the underlying system parameters are derived and the optimization problem is analyzed. In Section 4, the numerical results and performance evaluation are presented. Conclusion and ideas for further work are finally posed in Section 5.

In the tsj, the node senses the spectrum and makes a local decision about the presence or absence of the primary user based on its observations. Then the one bit local results are informed to the fusion center. The final decision is made at the fusion center by employing the OR rule, which means that the fusion center decides for the presence of primary user if one CRSN node reports as such.

2. SYSTEM MODEL

Denote Nj to be the sampling amount of the j-th node, where Nj∈[1, N ]. And fs is the sampling frequency. Each node has different Nj but same fs. Thus, the tsj of the j-th

In this paper, we consider a joint energy efficiency optimization of both spectrum sensing and data transmission, as shown in Fig.1. We define source sensing as the process of collecting various source information and delivering it to the Access Point. The main objective is realizing accurate acquisition of the source information. And the proposed dynamic censored spectrum sensing

node could be expressed as tsj = N j / f s .Thus, the tdj could be expressed as tdj = T - tsj .

135

Fig.3. Dynamic censored based spectrum sensing scheme. (a) The j-th node will keep sensing the spectrum if Tnj∈(λ1, λ2) until the sampling number reaches the maximum sensing sample number N; (b) The j-th node will stop sensing and declare H1 if Tnj ≥ λ2; (b) The j-th node will stop sensing and declare H0 if Tnj ≤ λ1.

To make local decisions about the presence or absence of a primary user, each node solves a binary hypothesis problem. Denote si to be the primary signal at time i with

Tnj can be defined as Tnj =

1 n 2 1 n z ij , where å xij = n å n i =1 i =1

z ij =xij 2 .

zero-mean and variance s s2 , ωij to be the additive white 2.1 Dynamic Censored based Cooperative Spectrum Sensing When the accumulated energy of the observation samples is calculated, the dynamic censored based spectrum sensing approach is employed at each CRSN node as shown in Fig.3. Each CSRN node can decide when to stop sensing, and the local results are sent to the fusion center only if they are deemed to be informative. Denote λ1 and λ2 to be the censoring thresholds of nodes, and the range (λ1, λ2) is called the censoring region. This way, the local decision rule in each node is as follows.

2

Gaussian noise with zero-mean and variance s w , hij to be the channel gain between primary user and the j-th CRSN node, and xij to be the i-th received sample at the j-th CRSN node. The underlying hypotheses are defined as H0 and H1 indicating the absence and presence of the primary user, respectively. Thus, the binary hypothesis testing problem is to determine H0 (i.e., xij = ωij ) and H1 (i.e., xij = hij si + ωij). Each CRSN node employs an energy detector which calculates the accumulated energy of samples. Note that, unlike the fixed size censoring spectrum sensing, each CRSN node collects a specific number of samples in this paper, and each CRSN node continues sensing until it reaches a decision while not passing a limit of N samples. Thus, the energy detector employed by each CRSN node

ìsend 1, declaring H1 , if Tnj ³ l2 and n Î [1, N ], ï ï if Tnj Î (l1 , l2 ) and n Î [1, N ), ïcontinue sensing, (1) í if Tnj Î (l1 , l2 ) and n = N , ïno decision, ï ï îsend 0, declaring H 0 , if Tnj £ l1 and n Î [1, N ].

Denote ρj to be the average censoring rate at the j-th CRSN node, and δ0j and δ1j to be the respective average censoring probability under H0 and H1. Depending on the prior knowledge about the respective prior probabilities of the hypothesis H0 and H1, π0=Pr(H0) and π1=Pr(H1), we have (2)

r j = p 0d 0 j + p 1d 1 j where

(

)

(3)

(

)

(4)

d 0 j = Pr T1 j Î (l1 , l2 ) , , TNj Î (l1 , l2 ) H 0 Fig,2. Slotted sensing and transmission scheme.

d 1 j = Pr T1 j Î (l1 , l2 ) , , TNj Î (l1 , l2 ) H1

136

Since the sampling amount number is not constant here, it should be taken into consideration. Denoting Nj to be the random variable representing the number of samples required to announce the state of the primary user.

1 æ ö æs G2 ö K ç ÷ ç ÷ ç ÷ ç ÷ ç ÷ Dø L è ÷ RG = log 2 ç 2 ç sm æ 1 1 ö÷ 2 ç ÷ ç1 - 2 ÷÷ ç K ç ç è D s G ø ÷÷ è ø

Thus, the average sample number N j for the j-th CRSN node, could be denoted as

(

(

)

N j = p 0 E N j H 0 + p 1 E N j H1

)

Since the effective transmission is only existed in the fraction of Pr(success). Hence, the channel rate should be higher than source rate:

(5)

where the E(Nj|H0) and E(Nj|H1) is the expectation of sample number of the hypothesis H0 and H1. And they can be expressed as (6) and (7). n

1 æ ö æs G2 ö K ç ÷ ç ÷ ç ÷ ç ÷ ç ÷ (10) RG L èDø ÷ = log 2 ç RC = 2 ç sm æ 1 1 ö÷ Pr( success ) 2 Pr( success ) ç ÷ ç1 ÷ çç K çè D s G2 ÷ø ÷÷ è ø

E N j H 0 = å éëêi ´ Pr T1 j Î (l1 , l2 ) , , Tn-1 j Î (l1 , l2 ) , Tnj ³ l2 H 0 ùúû (6) i =1

(

)

(

n

)

On the other hand, according to the Shannon Channel Capacity Theorem, the channel rate should meet the following condition:

ù E N j H1 = å é ê ëi ´ Pr T1 j Î (l1 , l2 ) , , Tn-1 j Î (l1 , l2 ) , Tnj ³ l2 H1 úû (7) i =1

(

)

(

)

The detailed mathematical analysis of the expectation and censoring probability will be showed in the next section.

æ P ö÷ RC £ W log 2 çç1 + ÷ è N 0W ø

Denote the energy consumed in one sample as εs, and the transmission energy of a decision bit as εd. The average energy consumption C j of spectrum sensing at the j-th CRSN node can be expressed as:

(

)

(11)

where W is the channel bandwidth, and N0 is the unilateral noise power spectral density. Thus, the average energy consumption required of each CRSN node to correctly delivering information source data is as below.

S

C Sj = N j ´e s + 1 - r j ´e d

(9)

(8)

C Dj = [Pr( success ) + Pr(collision)]tdj P = [Pr( success ) + Pr(collision)]tdj ´ (2

2.2 Application Data Transmission In this subsection, the connection between the average energy consumed in data transmission and the effective



transmission probability will be explored. Denote C Dj to

= [Pr( success) + Pr(collision)]tdj N 0W L æ ö çæ æçs G2 ö÷ K1 ö 2 Pr( success )W ÷ ÷ çç çç D ÷÷ ÷ ´çççç sè2 æ ø ö ÷÷ - 1÷÷ 1 ÷ m ç1 ççç 1- K ççè D - s G2 ÷÷ø ÷÷ ÷ çè ÷ ø è ø

RC - 1) N 0W W

(12)

be average energy consumption of application data transmission task at the j-th CRSN node, and Pr(success) to be the effective transmission probability available for CRSN node.

where the Pr(collision) is the collision probability, and tdj

For simplicity, a Gaussian source S with zero mean

is the average data transmission time which can be

and variance s

2 G

is considered in this paper. And the

expressed as

source generates symbols at a constant L symbols per second. The communication channel of interest is AWGN channel. Every CRSN node’s observation includes a

tdj = T - tsj . And

tsj

is the average

spectrum sensing time which can be calculated by tsj = N j / f s .

Gaussian noise μi with zero mean and equal variance s m2 .

Since the fs is a constant, if the sampling number N j

Thus, the Gaussian source rate RG can be expressed with a required distortion D as follows [17]:

increases, the time for spectrum sensing tsj follows, and

137

the time

tdj

According to [18], the probability density function (PDF) of the energy detector model Tnj in section II, follows a central chi-square distribution with 2n degrees of freedom under H0, and follows a non-central chi-square distribution with 2n degrees of freedom and a non-centrality parameter with 2γj under H1. Denote γj as the received signal to noise ratio (SNR) of the primary user measured at the j-th node under H1, and it is known or can be estimated.

for data transmission will decrease.

According to (10), if the effective transmission probability Pr(success) is low, the channel rate in the transmission slots has to be very high to complete the task of data transmission in a limited effectively transmission time slot.

2.2 Joint Energy Consumption Model If more time slot as well as energy is allocated for spectrum sensing, the sensing results about the status of primary user are more accurate, and the probability of collision will be reduced. However, the remaining time and the energy efficiency for data transmission will be reduced as well. Thus, for an energy constrained CRSN node, how to improve the energy efficiency and reduce the energy consumption is a most important critical.

Therefore, the average censoring probability with a sampling number of n under H0 and H1 can be expressed as below. n

d 0 j (n)=Õ [Q ( i ( i =1

n

d 1 j (n)=Õ[Q( i =1

From (8) and (12), the total energy consumption of the j-th CRSN node can be modeled as follow: C Tj = C Sj + C Dj

j

s.t. PFA £ a , PD ³ b

(15)

i l i l ( 1 - g j - 1)) - Q( ( 2 - g j - 1))] (16) 2g j + 1 s w2 2g j + 1 s w2

Where Q is the tail probability of the standard Gaussian distribution and can be expressed as Q ( x) =

(13)

1 2p

¥

ò x

( ) 2

exp - t2 dt . The (3) and (4) could be

viewed as an extreme case with a maximum sampling number N.

Thus, the energy efficient problem can be defined as a minimization of the maximum average energy consumption per CRSN node subject to a constraint on the global probabilities of false alarm and detection as follows, min max C Tj

l1 l - 1)) - Q ( i ( 22 - 1))] s w2 sw

Defining T0j=0, the local probability of false alarm at the j-th node, Pfj can be written as (17), whereas the local probability of detection, Pdj is obtained as in (18). N

(

Pfj = å Pr T1 j Î (l1 , l2 ) , , Tn -1 j Î (l1 , l2 ) , Tnj ³ l2 H 0

(14)

i =1

N

= å [d 0 j (n - 1) ´ Q( n ( n =1

where PFA and PD represent the global probabilities of false alarm and detection, respectively. And the α and β are pre-specified detection design parameters. The objective of this paper is to determine the optimal censoring λ1 and λ2 such that the maximum average energy consumption per node is minimized subject to the

N

l2 - 1))] s w2

(

Pdj = å Pr T1 j Î (l1 , l2 ) , , Tn -1 j Î (l1 , l2 ) , Tnj ³ l2 H1 i =1

n

l = å [d 1 j ( n - 1) ´ Q ( 22 - g j - 1)] g 2 + 1 s n =1 j w N

)

)

(17)

(18)

For the expectation of sample number E(Nj|H0) and E(Nj|H1), the (6) and (7) can be rewritten as below.

constraints PFA £ a and PD ³ b .

(

)

(

)

N

E N j H 0 =å[n ´d 0 j (n - 1) ´ Q ( n (

3. ANALYSIS AND PROBLEM FORMULATION In this section, analytical expression for both local and globe probability of false alarm and detection are extracted. And the expectation of sample number, the effective transmission probability and the collision transmission probability can be calculated as well.

n =1

N

E N j H1 =å [n ´d 1 j (n - 1) ´ Q n =1

l1 - 1))] s w2

(19)

l n ( 2 - g j - 1)] (20) 2g j + 1 s w2

Assumed that the OR rule is employed in fusion center, the global probability of false alarm is

138

K

PFA =1 - Õ (1 - Pfj )

considered for the numerical results. For all scenarios, we set the authorized primary user’s occupation rate to be 0.3, which means the authorized primary user is active with this probability. The probability of false alarm and detection constraints are assumed to be α=0.1 and β=0.9 as determined by the IEEE 802.15.4 standard for cognitive radios[19]. The energy consumed per sample in spectrum sensing is εs=0.1mW, and the energy consumed for decision bit is εd=1mW. The source is of unit variance, i.e. s G2 =1; the symbol rate of the source is L=1M baunds; the distortion is constrained to be 0.1. The bandwidth of the considered AWGN channel is W=5MHz. In some of the scenarios, for the sake of simplicity, it is assumed that all the sensors experience the same SNR. This way, it is easy to show how the main performance indicators including the optimal average energy consumption per sensor. Fig.4 depicts the average spectrum sensing energy consumption per sensor versus the number of sensors foe the OR rule. The SNR is assumed to be 0dB, the π0 can take a value of 0.2 or 0.8, and the maximum sampling number is N=10. It is shown that as the number of cooperating sensors increases, the energy consumed per sensor decreases and saturates. And for both high and low values of π0, the dynamic censored spectrum sensing scheme outperforms the fixed spectrum sensing scheme.

(21)

j =1

And the global probability of detection is K

PD =1 - Õ (1 - Pdj )

(22)

j =1

Hence, the effective transmission probability and the collision transmission probability can be calculated as: ìPr( success ) = p 0 (1 - PFA ) ï í ïPr(collision) = p 1 (1 - PD ) î

(23)

Only in the p 0 (1 - PFA ) throughput the time domain, the node can transmit data successfully. And in the

p 1 (1 - PD ) throughput the time domain, the node will collides with primary user. Having the analytical expression above, we can convert the solving problem of the optimal maximum average energy consumption per node to the research of the optimal λ1 and λ2. This way, the energy efficiency problem (14) can be rewritten as below. min max C Tj l1 , l2

j

s.t. PFA £ a , PD ³ b

(24)

Since the problem (24) is not concave, it is complex to solve it using two-dimensional exhaustive search approach. Therefore, for convenience of calculations and reaching a good solution in a reasonable time, we set λ1=0. For a given λ1, it is easy to show that the sensing upper

Fig.4 the average spectrum sensing energy consumption per sensor versus number of cognitive radios for the OR rule.

threshold PFA-1 (a ) £ l2 £ PD-1 (b ) , where PFA-1 and PD-1 are

Fig.5 consider a scenario where the number of sensors is K=5; the maximum sampling number is N=30; and the π0 can take a value of 0.2 or 0.8. The average spectrum sensing energy versus SNR is analyzed in this scenario for the OR rule. As shown in this picture, the average energy consumption variation for the fixed spectrum sensing scheme is almost same for all the considered SNRs, the dynamic censored spectrum sensing scheme’s average energy consumption per sensor reduces significantly as the SNR increases. Since the average sampling number dramatically decreases with the SNR increase. This shows that as the SNR increases, the proposed dynamic censored spectrum sensing scheme becomes more important.

the inverse functions defined over the α and β. Thus, deriving the optimal λ2 is then straightforward. As in section 4, censored truncated sequential sensing performs better than traditional spectrum sensing in terms of energy efficiency.

4. NUMERICAL RESULTS To validate the analysis of the proposed energy efficient dynamic censored spectrum sensing scheme, we present several numerical results. We use Matlab as our simulator, and a cognitive radio sensor network is

139

sensing scheme’s energy efficiency also increases compared to the fixed spectrum sensing scheme. Finally, Fig.7 shows the average total energy consumption of both spectrum sensing and data transmission per sensor versus the probability of effective transmission with different source SNR. It shows that there is a unique valley point in every curve, which corresponds to the optimal total energy. Any other transmission probability will result in a higher total energy consumption.

4. CONCLUSION In this paper, a dynamic spectrum censored based energy-efficient technique joint with source distortion has been presented. And the goal of this paper is to minimize the total energy consumption in spectrum sensing and data transmission. In order to achieve this goal, a joint energy efficiency optimization method for spectrum sensing and data transmission is investigated in this paper. Then a dynamic censored based cooperative spectrum sensing approach is employed to decide when to stop sensing. The distortion constrained probabilistic transmission method is utilized for jointly optimize the energy efficiency of spectrum sensing and data transmission. We modeled their energy consumption properly and jointly analyzed them. It is shown that significant energy efficiency improvement could be achieved in this paper.

Fig.5 the average spectrum sensing energy consumption per sensor versus SNR for the OR rule.

Fig.6 the average spectrum sensing energy consumption per sensor versus the number of samples for the OR rule.

Acknowledgment The authors would like to acknowledge that this work was partially supported by the National Natural Science Foundation of China (Grant No. 61379111, 61071096, 61379111 and 61073103) and Research Fund for the Doctoral Program of Higher Education of China (Grant No. 20110162110042).

References [1] A. Fehske, G. Fettweis, J. Malmodin, and G. Biczok, “The Global Footprint of Mobile Communications: The Ecological and Economic Perspective,” IEEE Communications Magazine, vol. 49, no. 8, pp. 55-62, 2011. [2] O. B. Akan, O. Karli, and O. Ergul, “Cognitive radio sensor networks,” IEEE Network, vol. 23, no. 4, pp. 34-40, 2009. [3] G. P. Joshi, S. Y. Nam, and S. W. Kim, “Cognitive Radio Wireless Sensor Networks: Applications, Challenges and Research Trends,” Sensors, vol. 13, no. 9, pp. 11196-11228, 2013. [4] E. Hossain, V. Bhargava, “Cognitive wireless communication networks,” Berlin: Springer, 2007. [5] G. Ganesan, Y. Li, “Cooperative spectrum sensing in

Fig.7 the average total energy consumption per sensor versus the probability of successful transmission under different source SNR.

Fig.6 shows the average spectrum sensing consumption per sensor versus the number of samples for the OR rule and for a network of 5 cognitive radio sensors where each senser experiences a different channel gain and thus a different SNR. Arranging the SNRs in a vector γ=[1dB, 2dB, 3dB, 4dB, 5dB], and the π0=0.5. by increasing the maximum number of samples and thus the total sensing energy, the dynamic censored spectrum

140

[6]

[7]

[8]

[9]

[10]

[11]

[12]

cognitive radio, part II: Multiuser networks,” IEEE Transactions on Wireless Communications, vol. 6, no. 6, pp. 2214-2222, 2007. I. F. Akyildiz, B. F. Lo, R. Balakrishnan, “Cooperative spectrum sensing in cognitive radio networks: A survey,” Physical Communication, vol. 4, no. 1, pp. 40-62, 2011. S. Maleki, A. Pandharipande, and G. Leus, “Energy-efficient distributed spectrum sensing for cognitive sensor networks,” IEEE Sensors Journal, vol. 11, no. 3, pp. 565-573, 2011. S. Maleki, G. Leus, “Censored truncated sequential spectrum sensing for cognitive radio networks,” IEEE Journal on Selected Areas in Communications, vol. 31, no. 3, pp. 364-378, 2013. S. Maleki, G. Leus, S. Chatzinotas, and B. Ottersten, “To AND or OR: How Shall the Fusion Center Rule in Energy-Constrained Cognitive Radio Networks?” IEEE International Conference on Communications (ICC), 2014.(in press) M. Najimi, A. Ebrahimzadeh, S. M. H. Andargoli, and A. Fallahi, “A Novel Sensing Nodes and Decision Node Selection Method for Energy Efficiency of Cooperative Spectrum Sensing in Cognitive Sensor Networks,” IEEE Sensor Journal, vol. 13, no.5, pp. 1610-1621, 2013. X. Hao, M. H. Cheung, V. W. Wong, and V. C. Leung, “A coalition formation game for energy-efficient cooperative spectrum sensing in cognitive radio networks with multiple channels,” IEEE in Global Telecommunications Conference (GLOBECOM), pp. 1-6, 2011. A. S. Zahmati, X. Fernando, A. Grami. “A Hybrid Spectrum Sensing Method for Cognitive Sensor Networks,” Wireless Personal Communications, vol.

[13]

[14]

[15]

[16]

[17]

[18]

[19]

141

74, no. 2, pp. 953-968, 2014. A. Jain, D. Gunduz, S. R. Kulkarni, H. V. Poor, and S. Verdu, “Energy efficient lossy transmission over sensor networks with feedback,” IEEE International Conference on Acoustics Speech and Signal Processing (ICASSP), pp. 5558-5561, 2010. A. Jain, D. Gunduz, S. R. Kulkarni, H. V. Poor, and S. Verdu, “Energy-Distortion Tradeoffs in Gaussian Joint Source-Channel Coding Problems,” IEEE Transaction on Information Theory, vol. 58, no. 5, pp. 3153-3168, 2012. H. Z. Zhang, Z. Y. Zhang, X. M. Chen, and R. Yin, “Energy efficient joint source and channel sensing in cognitive radio sensor networks,” IEEE International Conference on Communications (ICC), pp. 1-6, 2011. M. Naeem, K. Illanko, A. Karmokar, A. Anpalagan, M. Jaseemuddin, “Energy-Efficient Cognitive Radio Sensor Networks: Parametric and Convex Transformations,” Sensors, vol. 13, no. 8, pp. 11032-11050, 2013. Y. Oohama, “Rate-distortion theory for Gaussian multiterminal source coding systems with several side informations at the decoder,” IEEE Transactions on Information Theory, vol. 51, no. 7, pp. 2577-2593, 2005. S. M. Kay, “Fundamentals of statistical signal processing: detection theory,” Upper Saddle River, New Jersey: Prentice Hall Signal Processing Series 1998. C. R. Stevenson, C. Cordeiro, E. Sofer, and G. Ghouinard, “Functional requirements for the 802.22 WRAN standard,” IEEE Tech. Rep. 802.22-05/0007r46, Sept. 2005.

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Buffer Constraints Aware Data Gathering with Emergencies for Wireless Sensor and Actuator Networks Weirong Liu, Yun He, Shuo Li*, Fu Jiang, Jun Peng School of Information Science and Engineering, Central South University Changsha, Hunan, China

Abstract

According to the mobility of actuators, the mobile data gathering can be divided into two categories. The first is uncontrollable mobile data collection, which means the actuator either moves randomly or along a fixed track. A random walking algorithm for actuator is firstly proposed [3] . The obvious feature of this algorithm can reduce energy consumption and disordered motion can be improved for further study. Then a heuristic solution called earliest deadline first(EDF) algorithm is proposed to address these problems[4]. In the algorithm, the actuator path can be guided using two variables, but the collection delay is still unsolved. Based on this, nowadays Zhao et al. [5] employ multi-input multi-output and space division multiple access techniques to collect data from several nodes at the same time. Animals like whale, deer etc. are used as actuators to monitor their wild living conditions[7-8], As for the fixed track, Jea et al. restrict the mobile actuator to move along straight lines[6] and the application scenario is very limited. Finally, the authors provide a scheduled path for actuators to collect the information in fixed positions[9-10]. All these algorithms have the common features: high stability and reliability. However, the uncontrolled mobility restricts their range of applications severely. The other category is controllable mobile data collection. In this category, the actuator can move freely in the network and the trajectory can be planned for specific purposes. Based on this, this category can be divided into two subclasses: single-hop and multi-hop. In the first subclass, the actuator can collect data of sensor nodes within one-hop. Somasundara et al.[11] use mobile elements to avoid data loss due to buffer overflow. Then a tour planning algorithm is presented to achieve the balance of energy consumption[12], in which actuator have to visit every sensor node within one-hop and collect data from them directly. This method can save a lot of energy to some extent, but the insufferable collection delay reduce its performance severely. In the sencond subclass, the actuator collect data with multi-hop farwarding. This scheme can effectively reduce the collection delay. A moving path planning algorithm is proposed, in which some turning points are selected as the collecting point for actuators to avoid the obstacles and collect data efficiently[13]. Luo and Hubaux[14] set the actuator path as the perimeter of the sensing area and it is proved to be energy-efficient. Zhao et al. propose a data gathering

Using the mobility of the actuator to achieve data collection is a challenging problem in wireless sensor and actuator networks(WSANs). The actuator is usually required to collect the whole information within time delays. However, in practice not all the information need to be collected inmmediately except the emergencies. In this paper, a mobile data gathering algorithm with emergencies is proposed based on improved dynamic shortest path tree. In this algorithm, the improved dynamic shortest tree is built firstly. Then considering the transmission hops, polling points are choosed to gather data collected by the sensor nodes in releted region. Finally, the buffer contraits and emergencies are taken as key paremeters to select a subset of polling points as anchor points for the actuator. The optimal path for actuator is also constructed to reduce energy consumption considering the communication radius. Simulation results demonstrate that the proposed algorithm can greatly reduce packet loss and extend the lifetime of network.

Keywords: wireless sensor and actuator networks, buffer constraits, emergencies, mobile data gathering.

1. Introduction In recent years wireless sensor networks(WSNs) have attractted much attention in many potential fields, such us fire monitoring, weather alerting, military safety, health care etc.[1-2]. A Wireless sensor network consists of a large amount of low-cost, resource-limited sensor nodes, which are only allowed to exchange message with its neighbors. Each sensor in WSNs is required to collect environmental data and to upload those data to outside sink by multi-hop relay. Because the battery of sensor nodes are non-rechargeable and non-replaceable, energy saving becomes a key problem in wireless sensor networks. To address this problem, the actuator is introduced into WSNs to achieve mobile data collection because of their unlimited hardware resources and mobility. The WSNs with actuators are called wireless sensor and actuator networks(WSANs). A typical scenario of WSANs is that a mobile actuator moves to the location of each sensor node and collect data one by one to save energy.

142

We consider the network size D  R  R , in which three components are deployed: (1) some static wireless sensor nodes, donated as V={v1,v2,...,vn} ; (2) one or more data collectors, donated as κ=(1,...,k); (3)Sink node in the center, donated as B. The network is modeled as a directed graph G(V,E), and

scheme that joint considers the full utilization of concurent data uploading and tour length minimization[15]. But this would not fundamentally solve the fact that sensor nodes with multiple data forwarding consume more power than common sensor nodes. To address this issue, residual power are introduced into data gathering as a polling-points selecting parameter. Li et al. [16] propose a uniform energy consumption algorithm, where dynamic shortest path tree is used to construct network topology and the energy cosumption can be balanced effectively within time delay. All the algorithm above can achieve a great performance in collecting data within delay. However in practice, not all the information need to be collected inmmediately except for the emergencies, and the sensor buffer constrait is also needed to be considered. Therefore, in this paper, we propose a mobile data gathering algorithm with emergencies(BCA-DGE). In this algorithm, the polling points is selected based on improved dynamic shortest path tree firstly, and only the emergency polling points can be visited as anchor points by actuator within time delay. The buffer contraits and emergencies are taken as key emergency paremeters to select the anchor points for the actuator. Then a optimal tour for actuator is proposed, in which the communication radius is considered to reduce the collection delay for one anchor point. Thus more polling points can be visited so that the energy consumption can be optimazed. The rest of the paper is organized as follows:Section 2 outlines the system model and problem description. Section 3 introduces our proposed data gathering algorithm by three steps. Section 4 gives the evaluation for the performance of our algorithm. Finally Section 5 is the conclusion.

E  { vi , v j : vi , v j  V , i  j} is the set of all directed

links. The network is divided into several regions, and the polling points is defined as the center of these regions. Each sensor nodes transmit their data to affliated polling points via the directed links. The link is irreversible and thus there is no circles in the network. The model is described in Fig 1.

Sensor

Polling Point

Relay routing path

Sink

Mobile collector tour

Emergency field Communication radius

Fig.1 Illustration of data gathering with emergency and

buffer

constraits

To deploy our algorithm and compare our algorithm with other algorithms, we have to make some assumptions in WSANs. (1) All the static sensor node has fixed data buffer and same initial energy, which means data stored in each node is limited. On the contrary, the actuators has no constraits about it. Thus the sensor nodes have to transmit their data to actuators as fast as possible before the buffer is full. (2) Considering the relay hops and residual energy, some sensor nodes are chosen as polling points. Each polling points is responsible to collect data from other sensor nodes within certain relay hops. (3) According to the polling points, the netwok is divided into many regions. Each sensor node in these regions generate data with a constant speed v, and detect the emergency information with a probability s. (4) The actuator starts from the sink, traverses some of the selected poling points one by one and finally back to the sink. (5) The location of all the sensor nodes and the actuator can be known by using GPS devices. (6) The actuator can get their estimation value of buffer

2. System Formulation and Problem Description In this section, we first give system model of the proposed mobile data gathering algorithm and some assumptions about the model. Then the data gathering model is descripted as an integer linear programming problem, and the optimal objective is also given. 2.1 System Model In WSANs, mobile collector was introduced to collect data due to its ability to move. The thought of BCA-DGE is selecting a subset of nodes as polling points, and then finding a appropriate tour for mobile collector so that it can travel more sensor nodes as possible. In this paper, we focus on the selection of the poling points with consideration of buffer Constraints and energy savings, then an adaptative delay aware Data gathering algorithm is proposed to achieve the collection by using mobile actuators.

143

occupation rate when receiving the packets from the polling points. (7) The sink can get the short message about the location of emergencies and based on it establish the optimal tour for actuators.

Eu

 d

3. BCA-DGE Algorithm Design

2.2 Problem Formulation For the evaluation of data gathering algorithm, we briefly describe the data gathering problem as an optimization problem, called integer linear programming problem. The optimization objective is find a shortest path for actuator to visit the selected anchor points. Here is the problem formulation: Minimize



u ,vS { B}

Luv

In this section, The adaptative delay aware data gathering alglrithm with a single actuator is proposed. The algorithm aims to achieve two objectives, one is to find a subset of sensor nodes as the polling points by considering residual energy and buffer contraits, and then select suitable nodes of them as anchor points for the actuator. The other is to find the shortest tour for actuator to visit all the anchor points. As disscussed earlier, the selection algorithm can be realized through three steps: (1) Building the improved dynamic shortest path tree, (2) Finding a subset of sensor nodes as polling points. (3) Searching for anchor points and optimal actuator path

(1)

Subject to

y

iu

1

i  V

(2)

y

iu

 Iu

u  V

(3)

xiju  I u

i, j , u  V

(4)

xiju  0.5( yiu  y ju )nij

i, j , u  V

(5)



u  V

(6)

uV

iV

i , jV ,i  j

xiju 



iV ,i  u

xiu

3.1 Improved dynamic shortest tree In this section, we build the WSAN topology by using dynamic shortest path tree. In WSAN, the sensor nodes collect environment information and transmit them to sink through single-hop or multi-hop. To achieve this, the first step is to find a routing tree for each sensor node. So far some scholars have done some research on it and propose some methods. in [17], a minimum Steiner tree is used to minimize the total cost of network, but it is not used in practice due to its instability. Then in [15], the shortest path tree is proposed by considering the distance between neighbors. This method can build WSAN topology and meet the requirement ofdata transmission. All the methods above is static and not suitable for the dynamic change of energy. To solve this problem, in [16], S Li propose a dynamic shortest path tree, which consider the distance and residual energy of sensor nodes. Since the transmission energy is is proportional to their transmission distance, so we make some improvement on it. To build our WSAN topology, we first give some definitions of link cost Cuv between neighbors as follows:

In the above formulation, formula (1) represents optimal objective, which means the minimum of the actuator path. Constraits (2-4) means each sensor node is affiliated to one and only one polling points so that the each sensor node transmit its data to corresponding polling point once. Constraits (5-6) forbid the existence of relaying circles, which means only the sensor affiliated to the same polling point can relay packets and it is irreversible. The common variable used in this paper are described in Table 1. Table 1. Common variable in this paper

Variables D  R R

Value Network size,R is the network length

V=(v1,...,vn)

A set of sensor nodes in WSANs

Luv S yiu

Distance between anchor points u and v A set of anchor points If i is affiliated to u, yiu=1, otherwise yiu=0 If u is a polling point, Iu=1, otherwise Iu=0 If the data of I forwarded by j and rooted on u, xiju=1,otherwise xiju=0 if sensor node i is a neighbor of sensor node j, nij=1, otherwise nij=0 The link cost between u and v Distance between nodes u and v

Iu xiju nij Cuv Duv

Residual energy of sensor node u Probability of emergency Relay hops

Cuv  aDuv2 / ( Eu  Ev )

(7)

where Duv denotes the distance between neighbor nodes, Eu and Ev is the residual energy of sensor node u and v respectively, a is a constant. Then according the link cost Cuv, the dynamic shortest path tree can be built using iterative algorithm. The best benefit using improved dynamic shortest tree is to balance the energy consumption of WSAN. In the initial phase, each sensor node has same initial energy and the link cost is similar to the shortest path tree. So the intial topology of our algorithm is the same as the

144

shortest path tree. As time goes by, the residual energy of each sensor node begin to change, and the shortest path tree will be rebuilt so that the sensor nodes with less residual energy can reduce their power. That’s the reason that the dynamic shortest path tree is employed to build the network topology. The example of our algorithm is as follows: A 25

C

B

(a) initial phase Sensor

1

Polling Point

1

3

(a)

(b)

A 48 86

25

2

2 3

2

C95 B

1

2

3

1

(b) updating phase

3

(c) Sensor

Relay routing path

(d) Polling Point

Sink

Relay routing path

Fig.3 Selection of polling points. (a) Improved dynamic shortest tree.

Fig.2 Example for improved dynamic shortest tree (a)initial phase, each

(b) The first iteration. (c) The second iteration. (d) End of selection.

sensor node has the same energy (b) updating phase, with unbalance

A example for this process is shown in Fig.3. We set d=2. In Fig.3(a), the network topology can be built using improved dynamic shortest tree in 3.1. In Fig.3(b), In the first iteration, we found the sensor node 1 is the fartest leaf node, which means node 1 has the maximum hops far from sink. Then according the selection process of polling points, we search for the d=2-hop parent point for node 1, and we find sensor node 2. We choose sensor node 2 as the first polling points, and delete all the child nodes and corresponding links related to node 2. But we still remain node 2 as a candidate point. In Fig.3(c), in the sencond interation, we find node 2 still the fartest leaf node and it is already a polling point, so we search for d/2=1 hop parent point and find node 3. Then node 3 and its child points are deleted from T’, and they are affiliated to polling point 2. The rest can be done in the same manner. Finally we get all the polling points and divide the network into seven divisions.

energy consumption

As shown in Fig.2, the routing tree is constructed using improved dynamic shortest tree. The number in each link donates the link cost described in formula(7). In the initial phase, the data of sensor node C is forwarded by sensor node B, just like the shortest path tree. But after a while, sensor node B cosumed more energy than sensor A because it has more child-nodes and need to forward more data than sensor node B. According to Formula(7), the link cost CBC increase rapidly. When the link cost CBCDa, back to Step 2, otherwise go to Step 5. Step5: Exit. An example for anchor points selection and actuator path is given in Fig 4. Node 2, 5, 8, 9, 15 and 21 are polling points selected in Chapter 3.2. Their buffer occupation rate are 20%, 34%, 87%, 56%, 66% and 41% respectively. In addition, sensor node 14 affiliated to node 8 detects emergency things and need to notice the sink immediately. As shown in Fig.4(a), node 8 is selected as an anchor point in the first iteration. Then in the second iteration node 15 is selected as an anchor point due to highest buffer occupation rate. After calculation, the length of actuator path is shorter than the threshold, The

9

(c)

Sensor

(d)

Polling Point

Relay routing path

Sink

Emergency field

Mobile collector tour

Communication radius

Fig 4. Anchor point selection and optimal tour for actuator. (a)Iteration 1. (b) Iteration 2. (c) Interation 3. (d)Final results.

4. Performance Evaluation In this section, we conducted some simulations to evaluate the performance of our BCA-DGE algorithm. and we also compare them with another two existing data gathering algorithm. The first algorithm is shortest path tree based data gathering algorithm called SPT-DGA[15], in which each polling point is in charge to collect data from other sensor nodes within hop bounds. Another algorithm is the in-degree priority algorithm called IPA[16], where a collection delay constraits are considered to planning the actuator path and the energy can be saved by using dynamic shortest path tree. Some basic simulation parameters are described in Table 2. Table 2. The initial value of simulation parameters

Simulation parameter Network size Number of Nodes Communication Radius Packet size Actuator moving speed Transfer rate Sensor buffer size Instant a

value 400*400 m2 256 23 m 100 byte 1 m/s 250 Kbps 100 packets 1

4.1 Network lifetime In the simulation, we conducted a subset of simulations to evaluate the network lifetime of our SPT-DGA

146

algorithm. We set the network lifetime as the operation time span from the sensor node deployment to the loss of first node. Based on this , we get some simulation results. Figure 5 plots the total energy of SPT-DGA, IPA and our BCA-DGE algorithm. As shown in the figure 5, BCA-DGE always outperforms the SPT-DGA and IPA from the persperctive of total energy of WSAN. The WSAN using SPT-DGA and IPA continue working until 124 and 165 rounds resperctively, but BCA-DGE continue working until 214 rounds. That’s because the distribution of BCA-DGE is more uniform than IPA, and more energy-saving than SPT-DGA due to the use of dynamic shortest tree.

are collected within time delay. 800 BCA-DGE IPA SPT-DGA

700

Packet loss/round

600 500 400 300 200 100 0

100

Fig 7.

4000 BCA-DGE IPA SPT-DGA

3500

2000 1500 1000 500

BCA-DGE IPA SPT-DGA

300 250 200 150 100 50

60

80

Fig 5.

100

120 140 160 Time[round]

180

200

220

240 0

Network energy versus time Fig 8.

Figure 6 plots the network lifetime along with the increasement of sensor nodes density. When the network density is not enough, SPT-DGA, IPA and BCA-DGE are more or less the same. But it is obvious that BCA-DGE outperforms another two kinds of algorithm. Therefore, our BCA-DGE is very energy-effective in dense network.

1400

1000 800 600 400 200

Fig 6.

100

200

300 400 Number of Nodes

300 400 Number of Nodes

500

Emergency packet loss versus number of nodes

In this paper, we proposed a energy-efficient data gathering algorithm with emergencies in wireless sensor and actuator network. To reduce the loss of packets, especially the emergency packets, we consider sensor buffer as a paremeter to select the anchor points and transmit their data to actuator within buffer constraits. In this algorithm, the polling points is selected based on improved dynamic shortest path tree firstly, and only the emergency polling points can be visited as anchor points by actuator within time delay. Then a optimal tour for actuator is proposed, in which the communication radius is considered to reduce the collection delay for one anchor point. Thus more polling points can be visited so that the energy consumption can be maximized. Simulation results demonstrate that the proposed algorithm can greatly reduce packet loss and extend the lifetime of network.

1200

0

200

4. Conclusions

BCA-DGE IPA SPT-DGA

1600

100

Fig 8 plot the emergency packet loss along with the number of nodes. As the network density increases, we can know our proposed BCA-DGE algorithm is obviously better than another two algorithm in the form of emergency packets loss.

1800

Network Lifetime[rounds]

500

Packet loss versus number of nodes

350

2500

0 40

300 400 Number of Nodes

400

Emergency Packet loss/round

Network Energy[J]

3000

200

500

Network lifetime versus number of nodes

4.2 Packet loss In this section, we investigate the number of lost packet over numbers of sensor nodes, expecailly for the emergency packets. The numbers of sensor nodes are set from 20 to 600, and the simulation results are shown in Figure 7. SPT-DGA and IPA have the higher packet loss than BCA-DGE when the number of sensor nodes increases. That’s mainly because our BCA-DGE take buffer constraits into accout and the emergency packets

References

147

[1] I. F. Akyildiz, W. L Su, Y. Sankarasubramaniam, E. Cayirci. A survey on sensor networks[J]. IEEE Communications Magazine, 2002, 40(8), pp. 102-114. [2] C. Y. Chong, S. P. Kumar. Sensor networks: evolution, opportunities, and challenges[J]. Proceedings of the IEEE, 2003, 91(8), pp.1247,1256. [3] C. Rahul, S. S. Roy, S. Jain. Data mules: Modeling and analysis of a three-tier architecture for sparse sensor networks[J]. Ad Hoc Networks, 2003, 1(2), pp. 215-233. [4] Y. Y. Gu, D. Bozdag, E. Ekici. Mobile element based differentiated message delivery in wireless sensor networks[C]. International Symposium on World of Wireless on Mobile and Multimedia Networks, New York, 2006, pp. 83-92. [5] M. Zhao, Y. Y Yang. A framework for mobile data gathering with load balanced clustering and MIMO uploading[C]. IEEE INFOCOM, Shanghai, China, 2011, pp. 2759-2767. [6] D. Jea, A.A. Somasundara, M.B. Srivastava. Multiple controlled mobile elements for data collection in sensor networks[C]. Proc. IEEE/ACM Distributed Computing in Sensor Systems, June 2005. [7] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. Peh, D. Rubenstein. Energy-efficient computing for wildlife tracking: design tradeo ffs and early experiences with zebranet[C]. Proc. Int Conf. Architectural Support for Programming Languages and Operating System(ASPLOS) , 2002. [8] M. A. Batalin, M. Rahimi, Y. Yu, D. Liu, A. Kansal, G. S. Sukhatme,W. J. Kaiser, M. Hansen, G. J. Pottie, M. Srivastava and D. Estrin, "Call and response: experiments in sampling the environment," ACMSenSys, 2004. [9] R. Pon, M. A. Batalin, J. Gordon, A. Kansal, D. Liu, M. Rahimi, L.Shirachi, Y. Yu, M. Hansen, W. J. Kaiser, M. Srivastava, G. Sukhatmeand D. Estrin,

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

148

"Networked infomechanical systems: a mobile embedded networked sensor platform," ACM/IEEE IPSN, 2005. A. A. Somasundara, A. Ramamoorthy and M. B. Srivastava, "Mobile element scheduling with dynamic deadlines," IEEE Trans. Mobile Com-puting, vol. 6, no. 4, 2007. M. Ma and Y. Yang, "Data gathering in wireless sensor networks with mobile collectors," IEEE IPDPS, 2008 M. Ma and Y. Yang, "SenCar: an energy-efficient data gathering mech-anism for large-scale multihop sensor networks," IEEE Trans. Parallel and Distributed Systems, vol. 18, no. 10, Oct. 2007 J. Luo and J. P. Hubaux, "Joint mobility and routing for lifetime elongation in wireless sensor networks," IEEE INFOCOM, 2005 J. Luo and J. P. Hubaux, "Joint mobility and routing for lifetime elongation in wireless sensor networks," IEEE INFOCOM, 2005. M. Zhao, Y. Y. Yang. Bounded relay hop mobile data gathering in wireless sensor networks[J]. IEEE Transactions on Computers, 2012, 61(2), pp. 265-277. S. Li, J. Peng, W.R. Liu, Z.F. Zhu, K.C. Lin. A uniform Energy Consumptin Algorithm for wireless sensor and actuator networks based on dynamic polling point selection. Sensors. 2013. G. L. Xing, M. M Li, T. Wang. Efficient rendezvous algorithms for mobility-enabled wireless sensor networks[J]. IEEE Transactions on Mobile Computing, 2012, 11(1), pp. 47-60. Y. Yuan, Y. Peng, S. Li, W. Tang. Efficient heuristic algorithm for the mobile sink routing problem. Journal on Commnications 2011, 32, 107-117. Keller J M, Gray M R, Givens J A. A fuzzy k-nearest neighbor algorithm[J]. Systems, Man and Cybernetics, IEEE Transactions on, 1985 (4): 580-585.

The Joint International Conference of ITCA 2014 & ISCIIA 2014 Changsha, Hunan, China, 15-20 September 2014

Fast Differential Evolution Algorithm based Energy-Efficient Optimization for Cooperative Spectrum Sensing in Cognitive Radio Sensor Network Weirong Liu, Gaorong Qin, Fu Jiang*, Shuo Li, Jian He, Jun Peng* *School of Information Science and Engineering, Central South University Changsha, Hunan, China Email: [email protected], [email protected] Abstract

radio sensor network. Nevertheless, wireless communication, characterized by multi-path fading and adjacent channel interference, may be greatly influenced by surrounding environment. What’s more, as for the communication and processing resource-constraint of sensor nodes, the standalone spectrum sensing by using the traditional methods such as matched filtering, energy sensing and feature detection, is not reliable. The emerging cooperative spectrum sensing technology could overcome this shortcoming through the fusion of multiple cognitive radio sensor nodes’ spectrum sensing information. Furthermore, through cooperative spectrum sensing among cognitive radio nodes, the fusion center could gather the global information of the network so as to find the useful information[5-6], at the same time the node’s individual performance and the overall spectrum utilization may be improved. Various spectrum sensing technology have been proposed for improving the spectrum sensing performance in cognitive radio sensor network, for example, in earlier work [7], a consensus scheme between the secondary users has been introduced to design an optimal data exchange selection mechanism for improving the performance of predicting the primary users’ action. In [8], an efficient one-bit hard decision based three-phase spatio-temporal sensing algorithm, which is suitable for large-scale distributed wireless sensor network, is proposed by sensing vacant spectrum in both time and space domain. A parallel cooperative spectrum sensing mechanism is proposed to maximize the throughout of cognitive data. Nevertheless, research effort on improving the spectrum efficiency neglects the energy-constraint of cognitive wireless sensor nodes. The increasing energy consumption will reduce the lifetime of network nodes. A distributed spectrum-aware clustering mechanism is proposed in [9], which can reduce the energy consumption of information transmission between nodes by optimizing the clustering number. Based on the censoring scheme, a paper [10] proposed a cooperative spectrum sensing scheme, by designing a Takagi-Sugeno fuzzy inference system to make decision for the primary users’ spectrum occupancy. The data, which will be submitted from sensor nodes to fusion center, can be reduced significantly by the use of censoring scheme. The

In this paper, an energy-efficient optimization for cooperative spectrum sensing is investigated in cognitive radio sensor network. With sleeping schedule and censoring scheme combined, the energy-efficient cooperative spectrum sensing is formulated as a multi-objective optimization problem by analyzing the global sensing rate and false alarm rate. A fast multi-objective differential evolution algorithm is proposed to solve the multi-objective optimization problem, which takes advantage of the opposition-based learning for initializing the population and the tournament scheme in the mutation step. To accelerate the convergence rate and maintain the diversity, a dynamical adjustment scheme with a crossover parameter and a new population selection scheme is proposed. By considering a sensor network based on the communication protocol 802.15.4/ZigBee, a comprehensive performance evaluation is presented. The simulation results show that the proposed fast multi-objective differential evolution algorithm can not only reduce the average node energy consumption, but also improve the global rate of sensing.

Keywords: Cognitive Radio Sensor Networks, Sensing and fusion, Energy Efficiency, Spectrum Sensing.

1. INTRODUCTION Wireless Sensor Network (WSN), as an important access to information technology, is a basic method to realize the fusion of physical world and information world[1-2]. However, the traditional fixed spectrum allocation can not be able to meet the demand of the spectrum for wireless sensor network. Cognitive radio technology offers a natural solution to solve this problem. Incorporating cognitive radio technology in wireless sensor network yields the Cognitive Radio Sensor Network (CRSN). By dynamically changing its operating parameters, cognitive radio sensor nodes could sense the spectrum, determine the idle bands, and makes use of these available bands in an opportunistic manner, improving the overall spectrum utilization[3-4]. Spectrum sensing technology plays an important role in cognitive

149

local decision together by hard or soft fusion. Finally, the final decision about the spectrum utilization will be made by fusion center and sent to the sensor nodes.

methods above reduce the energy consumption of sensor node by reducing the information transmission. On this basis, some works [11-13] make further efforts to optimize the energy efficiency by integrating sleeping schedule. The majority of existing energy-efficient optimization for cooperative spectrum sensing minimize the energy consumption under the constraints of certain sensing performance. However, the decrease of energy consumption will degrade spectrum sensing performance correspondingly. Consequently, it will directly affect the accuracy of spectrum decision. In this paper, we consider the energy-efficient optimization for cooperative spectrum sensing in cognitive radio sensor network. Mathematically, we formulate the energy-efficient cooperative spectrum sensing as a multi-objective optimization problem by analyzing the global spectrum sensing rate and global false alarm rate. A fast multi-objective differential evolution algorithm is proposed to solve the multi-objective optimization problem. To accelerate the convergence rate and maintain the diversity, a dynamical adjustment scheme with a crossover parameter and a new population selection scheme is proposed. The rest of the paper is organized as follows. Section Ⅱ presents the system model. In Section Ⅲ , the energy-efficient cooperative spectrum sensing is formulated as a multi-objective optimization problem. Section Ⅳ provides a fast multi-objective differential evolution algorithm to solve the multi-objective optimization problem. A comprehensive performance evaluation of the algorithm is given in Section Ⅴ . Section Ⅵ is a brief conclusion of the paper.

Cognitive sensor node Primary user

Fusion center

Control channel

Base station

Fig.1 The network structure of cooperative spectrum sensing.

Spectrum sensing can be regarded as a binary decision between 0 and 1, where 0 denotes that the spectrum is available to secondary nodes and 1 represents otherwise. Sensor node senses the authorized spectrum under a binary hypothesis H0 and H1, where hypothesis H0 represents the absence of the primary signal and hypothesis H1 denotes the presence of the primary signal. Sensor node collects T0 samples during one sampling process. Therefore, the k th sample sensed by node i can be described as follows:

under H 0  ni [k ] ei [k ]    si [k ]  ni [k ] under H1

(1)

where k = 1,2,…,N, N is the number of cognitive radio sensor nodes. si[k] is the received signal form primary users. ni[k] is the white Gaussian noise signal with the variance σn2, and it is a certain signal. Sensor node senses the authorized spectrum by enenrgy sensing. The collected energy by the node i energy can be described as follows:

2. SYSTEM MODEL AND PROBLEM FORMULATION We consider a cognitive radio sensor network that shares a common frequency band with primary users. Of interest in this paper is to formulate the energy-efficient cooperative spectrum sensing as a multi-objective optimization problem. Firstly, we analyze the energy consumption and sensing performance. In this section, we present the models for cooperative spectrum sensing energy-efficient optimization in cognitive radio sensor network, including spectrum sensing model and global sensing model.

T0

Ei   xi2 [k ]

(2)

k 1

Each node implements the censoring scheme to make decision. The censoring threshold is λ1