The Sectored Antenna Array Indoor Positioning System with Neural Networks

Transcription

1 Automaton, Control and Intellgent Systems 2016; 4(2): do: /j.acs ISSN: (Prnt); ISSN: (Onlne) The Sectored Antenna Array Indoor Postonng System wth Neural Networks Chh-Yung Chen 1, Yu-Ju Chen 2, Ya-Chen Weng 1, Shen-Whan Chen 3, Rey-Chue Hwang 4, * 1 Department of Computer and Communcaton, Shu-Te Unversty, Kaohsung Cty, Tawan 2 Department of Informaton Management, Cheng Shu Unversty, Kaohsung Cty, Tawan 3 Department of Communcaton Engneerng, I-Shou Unversty, Kaohsung Cty, Tawan 4 Department of Electrcal Engneerng, I-Shou Unversty, Kaohsung Cty, Tawan Emal address: mkechen@stu.edu.tw (Chh-Yung Chen), yjchen@csu.edu.tw (Yu-Ju Chen), s @stu.edu.tw (Ya-Chen Weng), jasonchen@su.edu.tw (Shen-Whan Chen), rchwang@su.edu.tw (Rey-Chue Hwang) * Correspondng author To cte ths artcle: Chh-Yung Chen, Yu-Ju Chen, Ya-Chen Weng, Shen-Whan Chen, Rey-Chue Hwang. The Sectored Antenna Array Indoor Postonng System wth Neural Networks. Automaton, Control and Intellgent Systems. Vol. 4, No. 2, 2016, pp do: /j.acs Receved: February 22, 2016; Accepted: March 21, 2016; Publshed: March 25, 2016 Abstract: Ths paper presents a sectored antenna array ndoor postonng system (IPS) wth neural network (NN) technque. The hexagonal postonng staton s composed of sx prnted-crcut board Yag-Uda antennas and Zgbee modules. The values of receved sgnal strength (RSS) sensed by wreless sensors were used to be the nformaton for object s poston estmaton. Two NN models, ncludng NN wth back-propagaton (BP) learnng algorthm and probablstc NN (PNN), were appled to perform the postonng work for a comparson. In the experments, an 8x8 square meters ndoor scene was performed and 288 ponts and 440 ponts were expermented n ths area. The postonng results show that both NN models have the average error less than 0.7 meter. In other words, the proposed postonng system not only has the hgh postonng accuracy, but also has the potental n real applcaton. Keywords: Sectored Antenna, Indoor Postonng System (IPS), Neural Network (NN), Receved Sgnal Strength (RSS) 1. Introducton It s well known that IPS has become more and more popular n the object searchng due to the rapd developments of wreless communcaton technque and personal network [1-3]. IPS s used to provde the locaton nformaton of person and devce. It s a system of locaton based servce (LBS) through the ntegraton of the wreless communcaton and nformaton servces. The object s accurate poston could be determned by usng such a servce system. It has been used wdely n the varous applcatons, such as cargo management, patent montorng, publc gudng system, etc. Undoubtedly, IPS wll certanly play a sgnfcant role for the smart lfe of human bengs n the future. Generally, the structure of IPS could be dvded nto two parts,.e. the postonng algorthm and the sensng nfrastructure. The postonng algorthm s the method of determnng the object s locaton. So far, three algorthms, ncludng trangulaton, scene analyss and proxmty, are manly used for the object s poston estmaton [4-9]. The sensng nfrastructure s related to the wreless communcaton technology used for IPS. Nowadays, the varous wreless communcaton technologes such as wreless local area network (WLAN) [10-13], wreless sensor network (WSN) [14-15], rado frequency dentfcaton (RFID) [16-18], Bluetooth [19-20], Zgbee [21-22], etc. have been wldly used n the sensng technque of IPS. Each IPS has ts advantage and lmtaton n accordance wth the developed element s functon. In whch, many postonng computatonal algorthms used the values of RSS sensed from the known reference nodes to calculate the object s coordnate [23-24]. But, snce the nfluences of external factors such as the obstacle of hndrance, the nose dsturbance and the dffracton of electromagnetc wave, the postonng computatonal method s stll a challengng topc n the research of IPS applcaton.

2 Automaton, Control and Intellgent Systems 2016; 4(2): In recent years, due to the powerful learnng and adaptve capabltes, NN technque has been employed nto the postonng applcatons [25-31]. It s used to catch the nonlnear mappng between the coordnate of object and RSS sgnals. Through a tranng process, the well-traned NN model then can be used to estmate the object s locaton based on RSS measurements. In ths research, two NN models, ncludng NN wth BP learnng algorthm and PNN, were appled to perform the postonng work for a comparson. The detaled NN models wll be descrbed n the followng secton. In order to mprove the postonng accuracy, Cdronal et al [32] desgned a new swtched beam array antenna for wreless ndoor postonng applcaton. The antenna s ntended to augment a wreless devce operatng as the coordnator or base staton, and ts desgn s sutable for nstallaton on the celng of any large ndoor space [33-34]. Smlar to [32], ths paper presents a novel ndoor postonng scheme whch s composed of array antennas and Zgbee modules. The nformaton of sgnal angle and RSS are used to estmate the object s locaton. The whole paper s organzed as follows. The proposed ndoor postonng system s presented n Secton 2. Secton 3 descrbes the NN models for the postonng estmaton. Secton 4 presents the relevant experments and results. At last, a concluson s gven n Secton The Proposed Indoor Postonng System In ths research, Fgure 1 shows the developed IPS module whch conssts of two parts. One part s the ndoor postonng staton and the other part s the embedded postonng devce. The ndoor postonng staton s composed of a sectored antenna array, a mcrocontroller and sx Zgbee modules. The sectored antenna array has sx prnted-crcut board (PCB) Yag-Uda antennas wth hexagonal arrangement whch can provde 360 degrees coverage. Fgure 1. The proposed ndoor postonng system. Yag-Uda antenna s one of the most successful rado frequency drectonal antenna desgns. It has been used n a wde varety of applcatons that the antenna desgn needs gan and drectvty [35]. Fgure 2 shows the fgures of Yag-Uda antenna and ts radaton pattern. Fgure 2. Yag-Uda antenna and radaton pattern.

3 23 Chh-Yung Chen et al.: The Sectored Antenna Array Indoor Postonng System wth Neural Networks The embedded postonng devce s developed to perform the ndoor postonng functon. Through a tranng process, NN model could calculate the object s poston by usng RSS fngerprnt dataset obtaned from the staton. Then, the postonng result could be dsplayed on the screen of ARM-based system. The RSS based postonng technque estmates the poston from the sgnals of RSS vector whch can be obtaned by wreless sensors. The rado propagaton model wth postonng algorthm s a common way to determne the dstance between an object and the staton. Its equaton s generally expressed as ( 10 n log d A) RSS 1 p + (1) = 10 Where n p s the sgnal propagaton constant, d s the dstance between object and sensor and A s the object s RSS when the dstance s 1 meter. 3. NN and Modfed PNN In our studes, tradtonal NN model wth BP learnng algorthm and modfed PNN model were employed nto the postonng works for a comparson. Two NN models are brefly descrbed as follows NN Model A three-layered feed-forward fully connected NN s used n the studes. The sze of NN s whch means NN has 6 nput nodes, 13 hdden nodes and 2 output nodes. It structure s shown n Fgure 3. Two output nodes estmate the values of x and y axes of object s coordnate, respectvely. Step 1: Intalze all weghts, ω j to small random values (typcally between -0.5 to 0.5). Step 2: Present an nput pattern and specfy the desred output. Calculate output usng the presents ω j. Step 3: Fnd the error term, δ for all nodes. If D j, O j and H j denote the desred value of j th output node, the computed value of j th output node, and the computed value of j th hdden node, then the error terms of all nodes could be calculated by usng the followng equatons. The error of output layer node j: δ = D O ) O (1 O ) (2) j ( j j j j The error of hdden layer node j: δ j = H j( 1 H j ) ωjkδk (3) where k s over all nodes n the layer above node j. Step 4: Adjust weghts by ω ( n + 1) = ω ( n ) + ηδ u + ζ( ω ( n ) - ω ( n - 1)) (4) j j j Where (n+1), (n) and (n-1) ndcate the next, present, and prevous teraton numbers, respectvely. u s the th nput connected wth node j. η s the learnng rate. ζ s a momentum that effectvely flters out hgh-frequency varatons of the error surface. Step 5: Present next nput pattern and go back to step Modfed PNN The modfed PNN was ntalzed by Zaknch [25]. In ths research, PNN s appled to estmate the coordnate of object. The archtecture of modfed PNN s shown n Fgure 4. It conssts of one nput layer, one pattern layer, one summng layer and one output layer. The algorthm of modfed PNN s descrbed as follows. Let C be a set of class vectors.e. IPS tranng data, whch s gven by k j { c, y),( c, y),( c m, y )} C = (5) ( m j where m s the number of class vectors. c contans sx RSS sgnals sensed by antenna and y s the scalar output related to c. Here, the probablty densty functon (PDF) of modfed PNN s defned as Fgure 3. The NN structure. The error back-propagaton (BP) learnng algorthm s adopted by NN model. The learnng process s all tranng nputs are presented cyclcally untl all weghts of NN are stablzed. The major steps of BP learnng algorthm are summarzed as follows [36-37]. T ( x c ) ( x c) Φ( x, c, σ) = exp 2 (6) 2σ where σ s the smoothng parameter of Gaussan functon, x s the tranng vector for class n the nput space. Thus, the output ŷ.e. the coordnate of object can be obtaned by

4 Automaton, Control and Intellgent Systems 2016; 4(2): m yˆ (x ) = z y Φ( x, c,σ ) =1 m z (7) z Φ( x, c,σ ) =1 where s the number of x assocated wth c. Fgure 6. The llustrated fgure for 288 postons. Fgure 4. The archtecture of PNN. Fgure 7. The llustrated fgure for 440 postons. 4. Experments and Results In ths research, an 8x8 square meters ndoor feld as shown n Fgure 5 s used for the experments. In order to test the ndoor postonng system developed, 288 and 440 postons (features) wthn the ntervals of 0.5 meter and 0.4 meter were measured, respectvely. The features collected by IPS staton are RSS sgnals measured by wreless sgnal recever. Fgure 6 and Fgure 7 present the llustrated fgures for 288 and 440 test postons, respectvely. Fgure 5. The ndoor expermental feld The Experments by NN wth BP Learnng Algorthm Frstly, NN model wth BP learnng algorthm was used to do the object s poston estmaton. For 288 ponts experment, two data groups were collected randomly, each group has 288 data sets and each set ncludes the nformaton of pont s coordnate (x, y) and sx RSS sensed values (RSS1, RSS2,, RSS6). Frst data group was used for NN s tranng and second data group was used for test. Table 1 lsts the mean absolute errors (s) of postonal estmatons by usng NN model wth 10 dfferent learnng rates. From the results shown, the best performance s taken by NN model wth 0.1 learnng rate. The tranng and test are 48.5 cm and cm, respectvely. Smlarly, two data groups for 440 ponts were collected ether. The same experment was redone by NN model. The postonng results s also shown n Table 1. Agan, the best performance s taken by NN model wth 0.1 learnng rate. The tranng and test are cm and cm, respectvely. In ths part of experments, NN s vewed as the nonlnear regresson model for performng a nonlnear nput-output mappng. NN generates an approxmate functon to the tranng data.

5 25 Chh-Yung Chen et al.: The Sectored Antenna Array Indoor Postonng System wth Neural Networks Table 1. The statstc errors of 288 and 440 postonal estmatons by NN. Learnng rate 288 ponts 440 ponts Tranng Test Tranng Test Avg Fgure 8. The plot of 288 postonal estmatons The Experments by Modfed PNN In the experment by usng PNN model, the same data groups were performed by modfed PNN wth 10 dfferent σvalues. Table 2 presents the estmated postonng errors. From the results shown, PNN model wthσ=0.01 has the best estmaton. The estmated s of 288 ponts and 440 ponts could reach to 1.96 cm and 1.72 cm. Fgure 8 and Fgure 9 show the plots of 288 and 440 postonal estmatons. The symbols of crcle and dot are actual and estmated coordnates of the expermental data. Unlke prevous NN s experments, n ths part of research, the modfed PNN s vewed a classfer whch s used to estmate the object s poston n accordance wth the features (RSS sgnals) sensed. Compare the results of Table 1 wth Table 2, t s clearly found that the postonng accuracy performed by modfed PNN s much better than NN wth BP learnng rule. Table 2. The statstc errors of 288 and 440 postonal estmatons by modfed PNN. σ 288 ponts 440 ponts σ= σ= σ= σ= σ= σ= σ= σ= σ= σ= Avg Concluson Fgure 9. The plot of 440 postonal estmatons. Ths paper presents a sectored antenna array ndoor postonng system whch structure ncludes the postonng algorthm and the sensng nfrastructure. The nfrastructure s composed of sx prnted-crcut board Yag-Uda antennas and Zgbee modules whch are used to generate and sense the RSS sgnals. Two NN models are the postonng methods to perform the object s poston estmaton accordng to the sgnals of RSS sensed. In our studes, the postonng accuracy performed by modfed PNN model s much better than tradtonal NN. That means the classfer by modfed PNN has the outperformance than the regresson functon generated by tradtonal NN n ndoor postonng applcaton. However, such a concluson s gven by the envronment of RSS based postonng system must be stable and the sgnals of RSS have no serous problem caused by the effects of nterference, dffracton or reflecton. From the results shown n Table 1 and Table 2. It s able to be found that the estmated accuraces performed by modfed PNN model are hghly related to the values of σ. On the contrary, compare wth the modfed PNN model, the NN model wth BP learnng s more stable n performng the estmatons. The varances of estmatons are small. Thus, how to establsh a more excellent and stable ndoor postonng system s stll the future work we wll contnue.

6 Automaton, Control and Intellgent Systems 2016; 4(2): Acknowledgements Ths research was supported by the Mnstry of Scence and Technology, Tawan, ROC under Contracts No. MOST E , No. MOST E and No. MOST E References [1] Y. Y. Gu, A. Lo, I. Nemegeers, A survey of ndoor postonng systems for wreless personal networks, IEEE Communcatons Surveys & Tutorals, vol. 11, no. 1, pp , [2] H. Lu, H. Darab, P. Banerjee, J. Lu, Survey of wreless ndoor postonng technques and systems, IEEE Trans. on Systems, Man, and Cybernetcs, vol. 37, no. 6, pp , [3] G. W. Sh, Y. Mng, Survey of ndoor postonng systems based on ultra-wdeband (UWB) technology, Lecture Notes n Electrcal Engneerng, Wreless Communcatons, Networkng and Applcatons, Proceedngs of WCNA 2014, Vol. 348, pp , [4] B. Km, W. Bong, Y. C. Km, Indoor localzaton for W-F devces by cross-montorng AP and weghted trangulaton, In the Proceedngs of IEEE Consumer Communcatons and Networkng Conference (CCNC), NV, U.S.A., pp , [5] Y. Mo, Z. Z. Zhang, Y. Lu, G. Agha, A novel technque for human traffc based rado map updatng n W-F ndoor postonng systems, KSII Transactons on Internet and Informaton Systems, vol. 9, no. 5, pp , [6] X. F. Jang, C. J. Mke Lang, K. F. Chen, B. Zhang, J. Hsu, J. Lu, B. Cao, F. Zhao, Desgn and evaluaton of a wreless magnetc-based proxmty detecton platform for ndoor applcatons, In the Proceedngs of 11th ACM/IEEE Internatonal Conference on Informaton Processng n Sensor Networks (IPSN/SPOTS), Bejng, Chna, pp , [7] J. Hghtower, G. Borrello, Locaton sensng technques, Techncal Report UW CSE , Department of Computer Scence and Engneerng, Unversty of Washngton, [8] K. Kaemarungs, P. Krshnamurthy, Propertes of ndoor receved sgnal strength for WLAN locaton fngerprntng, In the Proceedngs of 1st Annual Internatonal Conference on Moble and Ubqutous Systems: Networkng and Servces (MobQutous 04), MA, USA, pp , [9] D. Focken, R. Stefelhagen, Towards vson-based 3-D people trackng n a smart room, In the Proceedngs of 4th IEEE Intl Conference on Multmodal Interfaces, PA, USA, pp , [10] A. Kotanen, M. Hannkanen, H. Leppakosk, T. D. Hamalanen, Postonng wth IEEE b wreless LAN, In the Proceedngs of 14th IEEE Proceedngs on Personal, Indoor and Moble Rado Communcatons, Bejng, Chna, pp , [11] Y. B. Xu, M. Zhou, L. Ma, Hybrd FCM/ANN ndoor locaton method n WLAN envronment, In the Proceedngs of IEEE Youth Conference on Informaton, Computng and Telecommuncatons, Bejng, Chna, pp , [12] V. Honkavrta, T. Perala, S. Al-Loytty, R. Pche, A comparatve survey of WLAN locaton fngerprntng methods, In the Proceedngs of 6th Workshop on Postonng, Navgaton and Communcaton (WPNC 09), Hannover, Germany, pp , [13] M. Y. Umar, K. V. Ramana, D. K. Yang, An enhanced K-Nearest Neghbor algorthm for ndoor postonng systems n a WLAN, 2014 IEEE Computers, Communcatons and Its Applcatons, pp , January 20, [14] K. F. S. Wong, I. W. Tsang, V. Cheung, S. H. G. Chan, J. T. Kwok, Poston estmaton for wreless sensor networks, In the Proceedngs of IEEE Global Telecommuncatons Conference, MO, USA, pp , [15] S. Aomumpa, K. Kondee, C. Prommak, K. Kaemarungs, Optmal placement of reference nodes for wreless ndoor postonng systems, 11th Internatonal Conference on Electrcal Engneerng, Electroncs, Computer, Telecommuncatons and Informaton Technology. Paper no , [16] P. Bahl V. N. Padmanabhan, RADAR: An n-buldng RF-based user locaton and trackng system, In the Proceedngs of INFOCOM 2000, Nneteenth Annual Jont Conference of the IEEE Computer and Communcatons Socetes, Tel Avv, Israel, pp , [17] H. D. Chon, S. Jun, H. Jung, S. W. An, Usng RFID for accurate postonng, Journal of Global Postonng Systems, vol. 3, pp [18] H. L. Dng, W. W. Y. Ng, P. P. K. Chan, D. L. Wu, X. L. Chen, D. S. Yeung, RFID ndoor postonng usng RBFNN wth L-GEM, In the Proceedngs of IEEE 2010 Internatonal Conference on Machne Learnng and Cybernetcs, Qngdao, Chna, pp , [19] A. K. M. M. Hossan, W. S. Soh, A comprehensve study of Bluetooth sgnal parameters for localzaton, In the Proceedngs of 18th Annual IEEE Internatonal Symposum on Personal, Indoor and Moble Rado Communcatons (PIMRC 07), Athens, Greece, pp. 1-5, [20] F. Subhan, H. Hasbullah, A. Rozyyev, S. T. Bakhsh, Indoor postonng n Bluetooth networks usng fngerprntng and lateraton approach, In the Proceedngs of 2011 Internatonal Conference on Informaton Scence and Applcatons (ICISA), Jeju Island, Korea, pp. 1-9, [21] W. P. Chen, X. F. Meng, A cooperatve localzaton scheme for Zgbee-based wreless sensor networks, In the Proceedngs of 14th IEEE Internatonal Conference on Networks, Sngapore, pp. 1-5, [22] G. Goncalo, S. Helena, Indoor locaton system usng ZgBee technology, In the Proceedngs of Thrd Internatonal Conference on Sensor Technologes and Applcatons, Athens/Glyfada, Greece, pp , [23] S. Merat, W. Almuhtad, Wreless network channel qualty estmaton nsde reactor buldng usng RSSI measurement of wreless sensor network, In the Proceedngs of Canadan Conference on Electrcal and Computer Engneerng, Calgary, AB, Canada, pp , 2009.

7 27 Chh-Yung Chen et al.: The Sectored Antenna Array Indoor Postonng System wth Neural Networks [24] H. C. Chen, Y. J. Chen, C. Y. Chen, S. M. T. Wang, J. P. Yang, R. C. Hwang, A new ndoor postonng technque based on neural network, Advanced Scence Letters, vol. 19, no. 7, pp , [25] A. Zaknch, Introducton to the modfed probablstc neural network for general sgnal processng applcatons, IEEE Transactons on Sgnal Processng, vol. 46, no. 7, pp , [26] R. C. Chen, Y. H. Ln, Usng ZgBee sensor network wth artfcal neural network for ndoor locaton, In the Proceedngs of Eghth Internatonal Conference on Natural Computaton, Chongqng, Chna, pp , [27] M. Altn, D. Brunell, E. Farella, L. Benn, Bluetooth ndoor localzaton wth multple neural networks, In the Proceedngs of 5th IEEE Internatonal Symposum on Wreless Pervasve Computng (ISWPC), Modena, Italy, pp , [28] Y. S. Ln, R. C. Chen, Y. C. Ln, An ndoor locaton dentfcaton system based on neural network and genetc algorthm, In the Proceedngs of 3rd Internatonal Conference on Awareness Scence and Technology (CAST), Dalan, Chna, pp , [29] H. Mohammad, A. F. Ozan, A. N. Al, P. Aveh, Neural network asssted dentfcaton of the absence of drect path n ndoor localzaton, In the Proceedngs of IEEE Global Telecommuncatons Conference, Washngton, DC, USA, pp , [30] S. H. Fang, T. N. Ln, Indoor locaton system based on dscrmnant-adaptve neural network n IEEE envronments, IEEE Transactons on Neural Networks, vol. 19, no. 11, pp , [31] C. Laoudas, D. G. Elades, P. Kempp, C. G. Panayotou, M. M. Polycarpou, Indoor localzaton usng neural networks wth locaton fngerprnts, Lecture Notes n Computer Scence - 19th Internatonal Conference on Artfcal Neural Networks, Lmassol, Cyprus, vol. 5769, no. 2, pp , [32] A. Cdronal, S. Maddo, G. Gorgett, G. Manes, Analyss and performance of a smart antenna for 2.45-GHz sngle-anchor ndoor postonng, IEEE Trans. Mcrow. Theory Tech., vol. 58, no. 1, pp , [33] C. H. Lm, Y. Wan, B. P. Ng, C. M. S. See, A real-tme ndoor WF localzaton system utlzng smart antennas, IEEE Transactons on Consumer Electroncs, vol. 53, no. 2, pp , [34] X. M. Qng, Z. N. Chen, T. S. P. See, Sectored antenna array for ndoor mono-staton UWB postonng applcatons, In the Proceedngs of 3rd European Conference on Antennas and Propagaton, Berln, Germany, pp , [35] Antennas and Propagaton. [36] A. Khotanzad, R. C. Hwang, A. Abaye, D. Maratukulam, An adaptve modular artfcal neural network: Hourly load forecaster and ts mplementaton at electrc utltes, IEEE Transactons on Power Systems, vol. 10, pp , [37] R. C. Hwang, P. T. Hsu, J. Cheng, C. Y. Chen, C. Y. Chang, H. C. Huang, The ndoor postonng technque based on neural networks, In the Proceedngs of IEEE Internatonal Conference on Sgnal Processng, Communcatons and Computng (ICSPCC 2011), X'an, Chna, pp , 2011.