1.0 INTRODUCTION 2.0 CELLULAR POSITIONING WITH DATABASE CORRELATION

Transcription

1 An Improved Cellular postonng technque based on Database Correlaton B D S Lakmal 1, S A D Das 2 Department of Electronc & Telecommuncaton Engneerng, Unversty of Moratuwa. { 1 shashka, 2 dleeka}@ent.mrt.ac.lk ABSTRACT The paper proposes an mproved network-based postonng technque based on database correlaton and presents the applcaton of the proposed technque to GSM. The fngerprnt database s created usng feld measurements and a novel and a more practcal method for gatherng fngerprnts s ntroduced. The authors propose a Weghted k-nearest Neghbor (WkNN) method for locaton estmaton and the tral results obtaned n urban and suburban envronments n Sr Lanka are presented. Comparson wth the Geometrcal method shows that the proposed method s a compettve alternatve for GSM locaton. 1.0 INTRODUCTION Postonng of Moble Statons (MSs) s a key problem n cellular networks. Ths s the problem of determnng the poston of a MS usng locaton senstve parameters. Inexpensve but accurate postonng systems are of great commercal potental due to the emergence of a multtude of locaton-based servces nvolvng a wde range of cvlan as well as mltary applcatons. These have great potental n areas such as emergency servces, personal securty, navgaton, toursm and entertanment. Provdng locaton-senstve servces such as ATM machnes, hosptals etc. are the other areas where the knowledge of user poston s mportant. Due to above mentoned and thousand more reasons there s no doubt n the mportance of achevng more accuracy n locaton estmaton. Currently the Global Postonng System (GPS) s a wdely used soluton for postonng. Even though GPS s farly accurate [1] there are some negatve ssues lke poor ndoor coverage, poor coverage n bad weather condtons or shaded areas and the necessty of beng equpped wth a GPS recever. On the other hand, cellular networks have provded excellent coverage n both outdoor and ndoor envronments and the cellular phone s becomng an everyday commodty among general publc. These ssues encourage the development of accurate cellular postonng technques. Cellular postonng technques can be categorzed nto Network based methods and Termnal based methods. In network based methods, trackng and evaluaton of the user locaton s carred out at the network sde and n termnal based methods the user devce performs the locaton estmaton tself. Some of these methods may or may not requre modfcatons to be done at the network sde or the termnal sde. The paper presents the research work carred out to come up wth a network based Improved Cellular Postonng Technque, whch can gve the coordnates of a moble staton locaton to a satsfyng accuracy. After comprehensvely studyng the currently used postonng methods, the authors have focused on developng a postonng technque based on database correlaton. The performance of the developed technque has been verfed, n urban and sub urban envronments through feld test trals. The rest of the paper s organzed as follows. In secton 2, the proposed based postonng technque s descrbed and secton 3 presents the encouragng tral results obtaned. Fnally, secton 4 concludes and suggests possble drectons for future research. 2.0 CELLULAR POSITIONING WITH DATABASE CORRELATION The database correlaton method () has been found n prevous studes to have promsng performance [2,3] n cellular postonng. In Database Correlaton, locaton-dependent network parameters are stored n the form of fngerprnts n a database constructed usng feld measurements or rado wave propagaton predcton tools. To estmate the locaton, a measured fngerprnt from a MS s correlated wth those n the database and the best matchng database fngerprnts are used for locaton estmaton. In general, ths nvolves three steps: 1. Obtanng fngerprnts 2. Database preparaton 3. Locaton estmaton 2.1 Obtanng Fngerprnts and Database Preparaton The stored sgnal nformaton samples n the database are called fngerprnts. Dependng on the partcular cellular system, the sgnal fngerprnts could nclude sgnal strength, sgnal tme delay, or even channel mpulse response. Also, t s possble to use measurements performed by the network as well as by the MS [2].

2 For the proposed method, the Receved Sgnal Strength (RSS) at the MS was selected as the fngerprnt varable and the fngerprnts were constructed usng an extensve set of measurements. Thus, a sngle fngerprnt n the database conssts of GPS coordnates of a locaton Receved sgnal strength (from servng base staton and other neghborng base statons) n that locaton Although a fngerprnt collectng process s llustrated n [2], the authors propose a novel and a more practcal method. Several drve tests were carred out and measurements were taken contnuously along roads usng an ntegrated GSM/GPS moble measurement unt developed by the authors. A sngle measurement consttutes the GPS coordnate of the pont at whch the measurement s taken and the RSS from all the hearable cells n that locaton. It was expected to obtan sgnal strengths from sx cells ncludng the servng cell and fve neghborng cells. Thus, the authors frst used a sngle measurement to form a fngerprnt as proposed n [2]. However, t was observed that the number of measurable neghborng cells s less than fve n most cases n the network envronment tested. Also the RSS vares sgnfcantly over tme at any locaton. Fg 1 llustrates the RSS varaton over tme at a partcular locaton for all the cells appeared n that locaton. Also t was observed that the set of cells that contrbute to consecutve measurements, along a road, were not always the same especally n urban areas. Fg 2 llustrates the RSS varaton of 10 consecutve measurements taken along a road for all the cells appeared n those 10 measurements. 90 RSS varaton over tme 85 RSS (dbm) cell A cell B cell C cell D cell E cell F cell G Cell_ID Fg 1: RSS varaton over tme for all the cells appeared RSS (dbm) RSS varaton along a road Measurement Fg 2: RSS varaton along a road for all the cells appeared cell A cell B cell C cell D cell E cell F cell G cell H cell I cell J

3 Accordng to Fg 1 t s evdent that a sngle measurement cannot be used to characterze the RSS at a locaton. But f fngerprnts are created usng number of statonary measurements at several locatons along a road, t wll not characterze the RSS n between the fngerprnt locatons. Because accordng to Fg 2 t s evdent that the RSS characterstcs vary sgnfcantly along a road. Therefore t was decded to take several consecutve measurements (along a road) together to form a sngle fngerprnt. In order to have a suffcent contrbuton from most of the hearable cells, ten consecutve measurements were taken to create a sngle fngerprnt by averagng the receved sgnal strengths (for each hearable cell). The locaton of the fngerprnt was consdered as the medan locaton of the correspondng ten ponts. Ths reduces the fngerprnt resoluton from a factor of 10 when compared wth that of [2]. Thus a Sldng wndow approach s proposed to reduce the dstance between two consecutve fngerprnts. For the mplementaton of the sldng wndow approach, only fve overlappng measurements were used n consecutve fngerprnts. Fg 3 llustrates ths sldng wndow approach of creatng fngerprnts. It was evdent from the results that ths approach resulted n better accuraces than usng separate sets of measurements for each fngerprnt. The fngerprnt resoluton can be further ncreased by ncreasng the number of overlappng measurements n the fngerprnts. Also the drve test was performed at a very low speed (less than 20km/h) to get a hgher resoluton and t was possble to obtan a fngerprnt resoluton less than 100m. Measurement Test route Fngerprnt 1 Fngerprnt 2 Fg 3: Obtanng fngerprnts-sldng Wndow method 2.2 Locaton Estmaton For locaton estmaton, the RSS of the locaton s to be measured and gven to the algorthm as the nput fngerprnt. Due to the fact that only one measurement cannot be used to characterze the RSS at a partcular locaton, several measurements should be obtaned at the test locaton and the averaged RSS of each hearable cell should be calculated and nput to the algorthm. The effect of the number of measurements taken at the requred locaton was studed and observed that better performance can be obtaned wth hgher number of measurements. Accordngly t was proposed to use 10 measurements at the locaton to be estmated. The authors examned the performance of both the Nearest Neghbor (NN) and the Weghted k Nearest Neghbor (WkNN) methods for locaton estmaton and proposed a locaton estmaton technque based on WkNN approach. The paper ntroduces two approaches for dentfyng the nearest neghbors. The frst approach (approach-i) s based on the dstance n sgnal space. The mplemented algorthm use the orgnal concept of Least Mean Square (LMS) approach publshed n [2]. But as n (1), the authors ntroduce a new cost functon based on Manhattan dstance along wth a new approach for the penalty term. d ( k) ( f g ( k)) ( f j lmax) w j ( lmax j k g ( k)) w (1) Where, d(k) = The dfference between the nput fngerprnt and the k th database fngerprnt. f = the averaged RSS of the th hearable cell n the nput fngerprnt g (k) = the averaged RSS of the same cell n the k th database fngerprnt f j = the averaged RSS of the j th hearable cell n the nput fngerprnt whch s not hearable n the k th database fngerprnt g k (k ) = the averaged RSS of the k th hearable cell n the database fngerprnt whch s not hearable n the nput fngerprnt l max = replace the mssng sgnal strength value w j, w k = correspondng weghts of the penalty terms The weghts of the penalty terms were calculated based on the contrbuton of the correspondng cell for the 10 measurements. The summaton s taken over the cells that are found n both fngerprnts. The database fngerprnts wth lower d(k) values are consdered as the Nearest neghbors of the locaton to be estmated. The second approach (approach-ii) s based on defnng a vald RSS range for each cell n each fngerprnt usng the averaged RSS and the standard devaton of the RSS of the correspondng cell. Here the dfference s calculated usng (2). c k f, g k, std k (2) k k

4 Where, c (k) = the total number of cells n the nput fngerprnt whch falls wthn the vald RSS range of correspondng cells n the k th database fngerprnt. f = the averaged RSS of the th hearable cell n the nput fngerprnt g (k) = the averaged RSS of the same cell n the k th database fngerprnt std (k) = the standard devaton of the same cell n the k th database fngerprnt δ (a,b,c) = a functon whch returns 1, f a falls wthn the regon b-c and b+c. Otherwse t returns 0. The summaton s taken over the cells that are common n both fngerprnts. The database fngerprnts wth hgher c(k) values are consdered as the Nearest neghbors of the locaton to be estmated. After dentfyng the nearest neghbors usng one of the above two approaches, the locaton s estmated usng the Weghted k-nearest neghbors (WkNN) method as below. p p w (3) p = estmated locaton p = locaton of the th nearest neghbor w = weght of the th nearest neghbor The results were tested wth three dfferent weghts. They are: w c c I. II. III. w w 1 d c d 1 d c d Also the performance of the WkNN estmaton was studed wth the geographcal K-Mean Clusterng algorthm wth K=2. However, a sgnfcant mprovement of the performance could not be observed wth clusterng. As proposed n [1] the number of searches was lmted based on the servng cell of the nput fngerprnt and each database fngerprnt. But ths crteron faled for some locatons n the urban envronment. Thus the authors came up wth a new soluton to fnd out the cell or cells whch have the hghest occurrence from all the hearable cells and to treat ths cell(s) also as a servng cell(s) when extractng the database fngerprnts. 3.0 RESULTS & DISCUSSION Feld trals were performed n two envronments: urban envronment (around Bambalaptya, n the heart of Colombo, Sr Lanka,) and a sub urban envronment (area around Unversty of Moratuwa). Encouragng results were obtaned for the tested envronments. Best results were obtaned by dentfyng the nearest neghbors usng approach-i. For urban envronment best results were obtaned wth weght I and for sub urban envronment best results were obtaned wth weght III. The performance of the WkNN estmaton was studed wth dfferent number of neghbors (.e. k) and best results were obtaned wth k=3 and k=4 for urban and sub urban envronments respectvely. Performance of the WkNN estmaton degrades wth hgher number of neghbors due to the affect of dstant neghbors for locaton estmaton. Fg. 4 and Fg. 5 show the CDF-wse comparson of the error for the best results obtaned for urban and sub urban areas respectvely. The results of the tral were compared wth the accuraces gven by the Geometrcal method, n whch the locaton s estmated by calculatng the dstance between the MS and BS for at least three MSs usng Hata model and applyng trlateraton [5,6]. For urban envronments, n 80% of the estmates, the error s less than 150m and for sub urban envronment t s less than 350m. The 80% error wth the Geometrcal method s 400m and 1250m n urban and sub urban envronments respectvely. Although s less accurate n sub urban envronment t shows excellent performance when compared wth the Geometrcal method. Table 1 and Table 2 summarze the results of the trals n urban and sub urban areas respectvely. Geometrcal Medan Error (m) % Error (m) Average Error (m) Standard Devaton Table 1: Summary of results for Urban envronment Geometrcal Medan Error (m) % Error (m) Average Error (m) Standard Devaton Table 2: Summary of results for Sub urban envronment

5 120% Error CDF comparson- Urban 100% 80% Percentage 60% 40% 20% 0% Geometrcal Error less than (m) Fg 4: CDF wse error comparson for urban area 120% Error CDF comparson- Sub urban 100% 80% Percentage 60% 40% 20% 0% Geometrcal Error less than (m) Fg 5: CDF wse error comparson for sub urban area 4.0 CONCLUSION The proposed based cellular postonng technque shows mproved postonng accuracy for both urban and sub urban envronments. Ths would be a better soluton for cellular postonng not only n urban envronments but also for less densely bult envronments where other postonng technques such as geometrcal method would fal to provde acceptable accuracy. Beng a network based method whch requres no modfcaton at the network sde or the termnal sde s an added advantage of the proposed method. 5.0 ACKNOWLEDGEMENT The authors are grateful to Dalog Telekom Ltd. for ther sponsorshp of ths research. 6.0 REFERENCES [1] Rchard Walter Klukas, A Super resoluton Based Cellular Postonng System Usng GPS Tme Synchronzaton, Department of Geomatcs Engneerng,The Unversty of Calgary, Alberta, December [2] Hekk Latnen, Jaakko Lähteenmäk, Tero Nordström, Database Correlaton Method for GSM Locaton, Vehcular Technology Conference (VTC), May [3] Paul Kempp, Database Correlaton Method for Multsystem Locaton, Master s thess, Department of Electrcal and Communcatons Engneerng, Helsnk Unversty of Technology, [4] Bnghao L, Yufe Wang, Hyung Keun Lee, Andrew Dempster,Chrs Rzos, A New Method of Yeldng a Database of Locaton Fngerprnts n WLAN, School of

6 Surveyng & Spatal Inf. Systems, Unversty of New South Wales, Sydney, Australa, October 2005 [5] Aleksandar Neskovc, Natasa Neskovc, George Paunovc, Modern Approaches In Modelng Of Moble rado Systems Propagaton Envronment, IEE Communcaton Servays, Thrd Quarter [6] Abhayawardhana, V.S., Wassell, I.J., Crosby, D., Sellars, M.P., Brown, M.G., Comparson of mprcal propagaton path loss models for fxed wreless access Systems, Vehcular Technology Conference, vol: 6, May -June 2005.