Implementing a Real-Time Locating System Based on Wireless Sensor Networks and Artificial Neural Networks to Mitigate the Multipath Effect

Transcription

1 14th Intenatonal Confeence on Infomaton Fuson Chcago, Illnos, USA, July 5-8, 2011 Implementng a Real-Tme Locatng System Based on Weless Senso Netwoks and Atfcal Neual Netwoks to Mtgate the Multpath Effect Dante I. Tapa, Rcado S. Alonso, Saa Rodguez, Fenando de la Peta, Juan M. Cochado Depatment of Compute Scence and Automaton Unvesty of Salamanca Salamanca, Span {dantetapa, alon, sg, fe, cochado}@usal.es Abstact - Weless Senso Netwoks compse an deal technology to develop Real-Tme Locatng Systems (RTLSs) amed at ndoo envonments, whee exstng global navgaton satellte systems do not wok coectly due to the blockage of the satellte sgnals. In ths egad, one of the man challenges s to deal wth the poblems that ase fom the effects of the popagaton of ado fequency waves, such as multpath. Ths pape pesents an nnovatve mathematcal model fo mpovng the accuacy of RTLSs, focusng on the mtgaton of the mulpath effect by usng Mult-Laye Pecepton Atfcal Neual Netwoks. The model s used to mplement a novel ndoo Real-Tme Locatng System based on Weless Senso Netwoks. Keywods: Weless senso netwoks, eal-tme locatng systems, multpath effect, atfcal neual netwoks. 1 Intoducton Whethe n home automaton, ndustal applcatons o smat hosptals, Weless Senso Netwoks (WSNs) ae used fo collectng useful nfomaton fo ntellgent envonments. In ths sense, Real-Tme Locatng Systems (RTLSs) compse one of the most nteestng applcatons fo WSNs. Among the most mpotant factos n the locatng pocess, we have the knds of sensos used and the technques appled fo the calculaton of the poston based on the nfomaton ecoveed by these sensos. Outdoo locatng s well-coveed by systems such as the cuent GPS (Global Postonng System) o the futue Galleo [1]. Howeve, ndoo locatng needs stll moe development, especally wth egad to accuacy and lowcost and effcent nfastuctues [2] [3]. Theefoe, t s necessay to develop Real-Tme Locatng Systems that allow pefomng effcent ndoo locatng n tems of pecson and optmzaton of esouces. In ths egad, the use of optmzed locatng technques allows obtanng moe accuate locatons usng even fewe sensos and wth less computatonal equements [2]. Among the weless technologes used to buld ndoo RTLSs, thee ae RFID (Rado Fequency IDentfcaton), Jave Bajo Faculty of Compute Scence Pontfcal Unvesty of Salamanca Salamanca, Span jbajope@upsa.es W-F (Weless Fdelty), UWB (Ulta-Wde Band) and the low-powe IEEE /ZgBee standad [4]. Nevetheless, ndependently of the technology used, t s necessay to establsh mathematcal models that allow detemnng the poston of a peson o object based on the sgnals ecoveed by the sensos nfastuctue. The poston can be calculated by means of dstnct locatng technques, such as sgnpost, fngepntng, tangulaton, tlateaton o multlateaton, among othes [4] [5]. Howeve, all of them must deal wth mpotant poblems when tyng to develop a pecse locatng system that uses WSNs n ts nfastuctue, especally fo ndoo envonments. Thee ae dffeent popagaton effects that affect the tansmsson and ecepton of the electomagnetc waves thoughout the weless senso nfastuctues used by these systems. Among these effects thee ae eflecton, scatteng, attenuaton and dffacton [6]. Thanks to the attenuaton effect, t s possble to estmate the dstance coveed by a wave between a tansmtte antenna and a eceve one [7]. Ths s vey useful to buld RTLSs based on these dstances, as those based on tlateaton [4]. Howeve, eflecton, dffacton and scatteng effects lead to othe poblems such as multpath. Ths way, the expected dstance coveed by a wave s deceased o even nceased due to the sum of waves eflected off the walls o the objects placed thoughout the envonment [7]. Specfcally, both outdoo and ndoo locatng systems based on the measuement of dstances between the sensos and the objects to be located can be affected by the gound eflecton effect [6], a knd of multpath popagaton effect. In ths sense, ths pape poposes a novel ndoo RTLS based on weless senso netwoks and a new mathematcal model that uses Atfcal Neual Netwoks (ANNs) as the man components to mtgate the gound eflecton effect and calculate the poston of the elements. Ths pape s stuctued as follows. Secton 2 explans the poblems that the gound eflecton effect ntoduces n ndoo RTLSs based on weless senso netwoks. Afte that, Secton 3 depcts a new poposal fo educng the gound eflecton effect by usng ANNs. Secton ISIF 756

2 descbes a novel ndoo Real-Tme Locatng System based on weless senso netwoks whee the new model has been appled to mtgate the gound eflecton effect, Fnally, Secton 5 pesents the conclusons obtaned and depcts the elated futue wok amed at mpovng the poposed method. 2 Poblem descpton Real-Tme Locatng Systems based on Weless Senso Netwoks can be seously affected by some effects elated to the electomagnetc waves popagaton, especally ndoos [2]. Some of these effects ae eflecton, scatteng o attenuaton, among othes. Such effects can povoke whch s known as multpath effect, and, moe specfcally, the gound eflecton effect [8]. 2.1 Real-Tme Locatng Systems The basc opeaton of a Real-Tme Locatng System s as follows. Fstly, t s necessay to deploy a netwok of efeence nodes thoughout the aea o envonment whee locatng wll be caed out. Such nodes ae usually called eades [4]. Some of these eades can move thoughout the montoed aea, actng as moble efeences. In addton, thee s a set of moble nodes, known as tags [3] [4], whch ae caed by the uses o assets to be located by the system. Each tag has a unque dentfe that unequvocally dentfes t n the system. Each of the tags sends a sgnal that contans ts dentfe n the system. Ths sgnal can be boadcasted peodcally o as a esponse to othe sgnals o exctatons tansmtted by the eades. In the latte case, the eades can be also called exctes [3]. The sgnals sent by the tags ae detected by the eades wthn the coveage aea. Thus, the eades can obtan the dentfe of the tag that sends each sgnal and also gathe some measuements of such a sgnal. These measuements gve nfomaton about the powe of the eceved sgnal (e.g., RSSI o Receved Sgnal Stength Indcaton), ts qualty (e.g., LQI o Lnk Qualty Indcato), ts Sgnal to Nose Rato (SNR) o the Angle of Aval (AoA) to the eade, among many othes. Fnally, the nfomaton (.e., the efeence measuements) fom all of the eades n the netwok s compled and pocessed to calculate the poston of each tag. Nevetheless, thee ae RTLSs n whch the poston s calculated by each tag though dffeent measuements fom sgnals tansmtted by dstnct efeence nodes and eceved at each tag (e.g., GPS). Howeve, n these knds of RTLSs only each use knows ts poston, unless use devces tansmt the own postons to anothe node though a cetan data communcaton channel, such as GPRS o UMTS [1]. Real-Tme Locatng Systems can be categozed by the knd of ts weless senso nfastuctue and by the locatng technques used to calculate the poston of the tags. Ths way, thee s a combnaton of seveal weless technologes, such as RFID, W-F, UWB and ZgBee, and also a wde ange of locatng technques that can be used fo detemnng the poston of the tags. Among the most wdely used locatng technques we have sgnpost, fngepntng, tangulaton, tlateaton o multlateaton [4] [5]. The set of the locatng technques that an RTLS ntegates s known as the locatng engne [4]. A wdespead technology used n Real-Tme Locatng Systems s Rado Fequency IDentfcaton (RFID) [3]. In ths case, the RFID eades act as exctes tansmttng contnuously a ado fequency sgnal that s collected by the RFID tags, whch n tun espond to the eades by sendng the dentfcaton numbes. When a tag passes though the eadng feld of a eade, t s sad that the tag s n that zone. Locatng systems based on Weless Fdelty (W-F) take advantage of W-F WLANs (Weless Local Aea Netwoks) wokng n the 2.4 and 5.8GHz ISM (Industal, Scentfc and Medcal) bands to calculate the postons of the moble devces (.e., tags) [9]. Howeve, locatng systems based on W-F pesent some poblems such as the ntefeences wth exstng data tansmssons and the hgh powe consumpton by the W- F tags. Ulta-Wde Band (UWB) s a technology whch has been ecently ntoduced fo developng these knds of systems. As t woks at hgh fequences (the band coves fom 3.1GHz to 10.6 GHz n the USA) [10], t allows to acheve vey accuate locaton estmatons. Howeve, at such fequences the electomagnetc waves suffe a geat attenuaton by objects (e.g., walls), so ts use on ndoo RTLS systems pesents mpotant poblems. ZgBee s anothe nteestng technology to buld RTLSs. The ZgBee standad s specally ntended to mplement Weless Senso Netwoks and, as W-F and Bluetooth, woks n the 2.4GHz ISM band, but can also wok on the MHz band. Dffeent locatng technques based on RSSI and LQI can be used on ZgBee WSNs (e.g., sgnpost o tlateaton). Moeove, ZgBee allows buldng netwoks of moe than 65,000 nodes n a mesh topology [4]. Among the locatng technques that make up the locatng engne we have the sgnpost technque, whch detemnes the aea on whch each tag n the envonment s located accodng to the closest eade to each tag [4]. On the contay, fngepntng s based on the pevous study of some measuements of the electomagnetc waves n each zone of the montoed envonment, thus estmatng n whch of such zones each tag s located [5]. On the one hand, tangulaton calculates the poston of each tag accodng to the angles of aval of the boadcasted sgnals between tags and eades [4]. On the othe hand, tlateaton calculates the poston of each tag fom the estmated dstance between each tag and a set of eades [1]. Fnally, multlateaton estmates the tags postons fom the Tme Dffeence Of Aval (TDOA) of the eceved boadcasts fom each tag to a set of eades [11]. The measuements used by these technques nclude the eceved powe of the sgnals (.e., RSSI), the qualty of the eceved sgnals (.e., LQI) o the angle of ncdence to the eceve antennas (.e., AoA). In an deal envonment, these measuements would be pefect, wth no eo o 757

3 nose, and the calculaton of tags postons would be exact. Nevetheless, n the eal wold, the electomagnetc waves ae nfluenced by popagaton effects such as eflecton, scatteng, attenuaton and dffacton. These effects can make the eades to eceve addtonal spuous sgnals that ae undesed copes of the man sgnal, whch makes up the multpath effect. When the gound s the man esponsble of waves eflectons, multpath can be modeled as the gound eflecton effect. 2.2 The gound eflecton effect The detaled effects of phenomena as attenuaton and eflecton n the popagaton of electomagnetc waves can be calculated by solvng Maxwell's equatons wth some bounday condtons that model the physcal chaactestcs of each object o medum nvolved [6]. As ths calculaton can be vey complex o the physcal chaactestcs of each object can be even unknown, thee ae some appoxmatons to model sgnal popagaton. One of these appoxmatons s the ay-tacng technque that smplfes electomagnetc wavefonts to smple patcles. Physcally, each wavefont s the locus of spatal ponts pesentng the same phase fo a cetan electomagnetc wave. In the ay-tacng technque, each wavefont s consdeed to be a patcle tavelng fom the tansmtte to the eceve antennas [6]. The popagaton of electomagnetc waves between antennas s a well studed physcal phenomenon. Let us consde that we have two antennas coectly algned and polazed between them (whose gans ae G T fo the tansmttng antenna and G R fo the ecevng antenna). Let us also suppose that such antennas ae sepaated a cetan d dstance between them n the fee space. Fnally, let us consde that one of these antennas, the tansmtte one, s tansmttng an electomagnetc wave to the othe antenna, the eceve one. When tansmttng a monochomatc electomagnetc wave (o an enough naow-band wave to assume a unque λ wavelength n the tansmsson medum) though the fee space between two antennas coectly algned and polazed, the eceved powe s gven by the Fs tansmsson equaton [7]: P R λ 1 = PT GT GR 4π d whee P R s the powe avalable fom the ecevng antenna, P T s the powe suppled to the tansmttng antenna, and n s the path loss exponent, that s expementally calculated (e.g., n the fee space n = 2.0). Howeve, n the eal lfe, an electomagnetc wave tansmtted by a cetan weless souce wll be eflected, dffacted o scatteed by the multple objects placed thoughout the envonment. Ths way, the antenna of the destnaton node wll eceve undesed copes of the tansmtted sgnal. Even wose, these addtonal sgnals wll be possbly delayed n tme and shfted n both fequency and phase. When a sngle gound eflecton n (1) effect pedomnates n the multpath effect, a two-ay model can be used. In ths model, t s consdeed that a adofequency sgnal s tansmtted though the fee space fom a tansmtte antenna to a eceve antenna. The dstance of the bases of tansmtte and eceve s d. It s supposed that the gound s a pefect nfnte flat plan. Moeove, the heght of the tansmtte antenna ove the gound s h T, wheeas the heght of the eceve antenna ove the gound s h R. Ths way, the total enegy of the sgnal n the eceve antenna s the sum of the enegy of the dectly tansmtted wave between both antennas and the enegy of one wave tansmtted fom the tansmtte antenna and eflected off the gound. Ths eflected wave comes n contact wth the gound wth a θ ncdent angle and t s eflected wth a θ eflecton angle (by the law of eflecton, θ = θ ). Due to the dffeence of the phases of dectly tansmtted and eflected waves, they wll be constuctvely o destuctvely added n the eceve antenna. As both the tansmsson medum (typcally the a) and the gound can be consdeed as delectc meda, a poton of the ncdent wave on the gound s eflected by the juncton between them and the est of the enegy passes though ths juncton. Ths way, consdeng both meda as non-conductve and beng the tansmsson medum the fee space and the gound a delectc whose elatve pemttvty s ε and elatve pemeablty μ = 1, the Fesnel eflecton coeffcents fo hozontal and vetcal polazed sgnals ae gven by: Γ Γ H V snθ = snθ + ε snθ = ε snθ + 2 ε cos θ 2 ε cos θ 2 ε cos θ 2 ε cos θ Ths way, Fgue 1 shows how gound eflecton effect nfluences the tansmsson ange between a tansmtte and a eceve n both the 2.4GHz (typcal band used n W-F, Bluetooth and ZgBee) and the 868MHz (used by ZgBee n Euope) bands nsde an offce wth soft pattons (n = 2.6, wth a standad devaton σ = 14.1, and a pemttvty ε = 18 fo the gound), whee tansmtte and eceve ae espectvely sted at 2.5m and 1.1m heght ove the gound. These condtons ae typcal fo ndoo RTLS usng WSNs. Ths fgue shows how, due to gound eflecton effect, the eal eceved waves have moe losses than deally eceved waves. Even wose, the eceved powe does not vay monotoncally wth dstance. Ths means that, fo cetan local dstances, we can have a hghe eceved powe even f the eceve s sted futhe away fom tansmtte and vce vesa. Futhemoe, f we want to guess the dstance between a tansmtte and a eceve fom the eceved powe, we can have moe than one eceved powe that could cause such a dstance. As can be seen n the fgue, the gound eflecton effect s geate fo hghe fequences. (2) 758

4 has aleady been taned usng standad locaton algothms. Fg. 1. Vaaton of the estmated dstance egadng the eceved powe at the eceve antenna usng a two-ay model to consde the gound eflecton effect. 2.3 Related appoaches Thee ae some elated appoaches focused on the study o the mtgaton of the multpath o the gound eflecton effect. Xe et al. [12] pesent a multpath mtgaton algothm fo a spead spectum ado-fequency system usng a fequency hoppng technque and consdeng a two-ay envonment. Howeve, ts ntegaton wth exstng RTLSs s had as t s based on the fequency hoppng technque and mples to modfy the weless technologes used n the systems. Eu Seok Km et al. [8] study the angng pefomance though a two-ay multpath smulated model. Ths eseach uses IEEE a as weless technology ntended fo ndoo locatng applcatons. Nevetheless, t s only a smulaton wth analyss poposals and does not popose a eal way to mtgate the gound eflecton effect. Thee ae othe appoaches that use Atfcal Neual Netwoks to pefomng locatng technques. Salcc and Chan [13] popose the use of an ANN to estmate moble GSM phone postonng. Ths eseach ncludes the compason of the ANN wth a functon appoxmaton model. Neguzan et al. [2] pesent a method fo ndoo moble staton locaton usng eceved sgnal stength fngepntng measuements appled to an ANN. 3 An nnovatve locatng model The nnovatve mathematcal model poposed n ths pape takes RSSI measuements as nputs and makes an estmaton of the poston of tags based on these RSSI measuements. RSSI has been selected because t s one of the most common measues used n cuent ndoo RTLSs and t s also easy to obtan n most weless technologes [4]. At a fst stage, the model establshes the most pobable dstance of each tag based on the RSSI levels by mtgatng the gound eflecton effect. In the second stage, the geneated data s used by the model to tan an MLP (Mult-Laye Pecepton) neual netwok [14] whch fnally estmates the postons of each tag afte the system 3.1 Mtgatng the gound eflecton effect In deal condtons, the modelng of the elatonshp between RSSI levels and dstances between antennas has a decayng exponental shape. Nevetheless, as shown n Fgue 1, when gound eflecton effect s taken nto account, the pocess of appoxmaton of the elatonshp between the RSSI levels and the dstances between antennas s complex and poblematc. Theefoe, t s necessay to use othe models that allow consdeng the gound eflecton effect n ode to obtan a elable estmaton of the dstances between tags and eades. Cuently thee s a wde ange of models fo functon appoxmaton. Among the most wdely used, we have the egesson models. Regesson models ae based on the geneaton of mathematcal functons that mnmze a cetan eo functon accodng to the tanng data. Usually, the eo functon utlzed s the least squae eo functon. Thee ae dffeent egesson models egadng the exstng elatonshp between the vaables. Ths way, we have the lnea egesson, exponental egesson o the logathmc egesson models, among many othes. An altenatve to these egesson models s the Suppot Vecto Regesson (SVR) [15] [16]. The SVR method comes fom a set of elated methods whch lean fom data known as Suppot Vecto Machne (SVM). SVM s a supevsed leanng technque that s appled to the classfcaton and egesson of dffeent elements. SVM facltates wokng wth data that cannot be adjusted to lnea models [15], ntally conceved to obtan classfcatons n lnea sepaable poblems. Othe egesson methods appled when the dstbuton of data and the elatonshps ae unknown ae supevsed leanng neual netwoks. These knds of ANNs ae appled n foecastng poblems as, fo nstance, electc powe consumpton [17], gas consumpton, o ol slck foecast [19]. In the wok of Kalogou [17] t s pesented a complete evew of case studes whee these atfcal neual netwoks have been appled. Among the supevsed leanng netwoks thee ae dffeent altenatves fo the leanng pocess. Ths way, we have eo-coectng leanng [18], delta ule o least squaes leanng [18], genealzed ule o eo backpopagaton algothm (genealzed delta ule) [18], enfocement leanng, o stochastc leanng. Among supevsed leanng netwoks we have the Mult-Laye Pecepton (MLP) o the adal bass functon (RBF) netwoks [20]. Atfcal Neual Netwoks allow wokng wth tme sees. The use of tme sees facltates the foecast f t s not possble to make estmatons of non-ndependent values wth consecutve samples. Ths way, t s povded a moe ealstc foecast of values. Indeed, ths s a fundamental featue fo the foecast of dstances fom the RSSI levels, thus mtgatng the gound eflecton effect. Ths s because the gound eflecton effect manly occus nsde cetan anges of the dstances. 759

5 As shown n Fgue 1, the gound eflecton effect egsteed n the sgnals vaes egadng the RSSI level. As can be seen n such a fgue, fo a cetan ange of RSSI values, thee ae fluctuatons n the dstance values egadng the RSSI levels. Thus, a cetan RSSI value can mean dstnct dstances. In ode to model the gound eflecton effect we utlze tme sees appled to Atfcal Neual Netwoks. Specfcally, we use a Mult-Laye Pecepton, as mentoned befoe. Atfcal Neual Netwoks allow foecastng a value accodng to the eceved hstocal values. Theefoe, n ths wok the neual netwok s povded as nputs both the cuent detected RSSI value and the RSSI values detected n pevous tme nstants. Ths s the way we ntend to mtgate the gound eflecton effect. The neual netwok s made up of n nput neuons, beng n the tme nstants taken nto account: t, t - 1,..., t - (n - 1). The ntemedate laye of the neual netwok s confgued followng the Kolmogoov theoem [21] and choosng 2n + 1 neuons. In ode to mpove the foecast of the tme sees t was opted to ncopoate the RSSI levels povded by othe eades nto the neual netwok. Ths way, the dstances foecastng s done usng a subset of the deployed eades n the system smultaneously. The achtectue of the neual netwok s depcted n the Fgue 2. Ths neual netwok has k nput goups wth n neuons each of them. These n neuons coespond wth the n values of the tme sees. Lkewse, the k goups coespond wth the numbe of eades that ae consdeed fo the dstance estmaton. Ths numbe of eades s set n advance, thus selectng the eades wth hghest measued RSSI levels fom the tag. The ntemedate laye s made up of 2(k + n) + 1 neuons, wheeas output laye s fomed by k neuons (.e., a neuon pe each eade). The goups of nput neuons ae odeed accodng to the cuent RSSI level fom hghest to lowest. Theefoe, the fst output of the neuons s assocated to the eade that eceved the hghest RSSI level and so on. 3.2 Estmatng the locaton of the tags Ou poposed model captues data fom the estmaton of the postons by the tlateaton algothm. It stoes these n a memoy to subsequently use to cay out the tanng of an MLP (Mult-Laye Pecepton). Ths way, the neual netwok allows us to make the fastest estmatons and s moe esponsve to vaatons n the dstances esultng fom the eflectons of the waves emtted. Input data n the neual netwok coesponds wth the dstances calculated by means of the MLP shown n Fgue 2 fom a pe-fxed numbe of eades and the poston of the eades. These eades ae selected accodng to the lowest dstances they have to the tag. Output has two coodnates, one fo each plane coodnate. The numbe of neuons n the hdden laye s 2n + 1, whee n s the numbe of neuons n the nput laye. Fnally, thee s one neuon n the output laye. The actvaton functon selected fo the dffeent layes has been the sgmod. Futhemoe, the neuons extng fom the hdden laye of the neual netwok contan sgmodal neuons. Netwok tanng s caed out though the eo backpopagaton algothm [14]. Fg. 2. Stuctue of the Mult-Laye Pecepton used n the tanng stage fo the mtgaton of the gound eflecton effect usng multple eades. The ANN contans n nputs fo each of the k eades and k outputs. In the poposed model, the MLP used fo estmatng the tags postons has to be taned n each envonment n ode to get the hghest accuacy n the locatng pocess. 4 Implementng a new ndoo RTLS In ode to test the pefomance of ths model nto an ndoo envonment, t has been developed and deployed n-coe Polas, a new ndoo RTLS based on ZgBee WSNs and the nnovatve locatng technque poposed n ths pape. The bass WSN nfastuctue of the system s made up of seveal ZgBee nodes (.e., eades and tags). As mentoned befoe, the ZgBee standad s specally ntended to mplement WSNs and allows opeatng n the fequency ange belongng to the ado band known as Industal, Scentfc and Medcal (ISM), specfcally n the 868MHz band n Euope, the 915MHz n the USA and the 2.4GHz n almost all ove the wold. The undelyng IEEE standad s desgned to wok wth lowpowe and lmted computatonal esouces. The ZgBee standad allows moe than 65,000 nodes to be connected n a sta, tee o mesh topology. These featues make ZgBee and deal suppotng weless technology fo buldng ndoo Real-Tme Locatng Systems. The possblty of wokng wth low-powe nodes that do not need lage computatonal esouces allows desgnes to educe hadwae costs when mplementng the systems. In addton, these knds of low-powe nodes can each battey lfe-tme even of seveal yeas, egadng the tansmsson ange (tansmtted powe), the tme esoluton and the accuacy of the system. ZgBee-based Real-Tme Locatng Systems can use dffeent locatng 760

6 technques n ode to estmate the postons of the tags n the envonment. module can access to a emote database n ode to obtan nfomaton about the uses and egste hstocal data, such as alets and locaton tackng. Fg. 3. n-coe Sus B/D devces used by the n-coe Polas RTLS. In the poposed system, each ZgBee node ncludes an 8-bt RISC (Atmel ATmega 1281) mcocontolle wth 8KB of RAM, 4KB of EEPROM and 128KB of Flash memoy and an IEEE /ZgBee tansceve (Atmel AT86RF230). These devces, called n-coe Sus B and Sus D ae shown n Fgue 3. They both have 2.4GHz and 868/915MHz vesons and nclude a USB pot to chage the battey o supply them wth powe [22]. Lkewse, the USB pot can be used to update the fmwae of the devces and confgue the paametes fom a compute unnng a specal applcaton ntended to t. On the one hand, n-coe Sus B devces ae ntended to be used wth an ntenal battey and nclude two geneal-pupose buttons. On the othe hand, n-coe Sus D devces ae amed at beng used as fxed ZgBee outes usng the mans powe supply though a USB adapto. In the n-coe Polas RTLS, n-coe Sus B devces ae used as tags, whle n-coe Sus D devces ae used as eades. Ths way, n-coe Sus B devces ae caed by uses and objects to be located, wheeas n-coe Sus D devces ae placed at celngs and walls n ode to detect the tags. In Fgue 4 t can be seen the basc achtectue of the n- Coe Polas Real-Tme Locatng System. The kenel of the system s a compute that s connected to a ZgBee netwok fomed by n-coe devces. That s, the compute s connected to an n-coe Sus D devce though ts USB pot. Such a devce acts as coodnato of the ZgBee netwok. The compute uns a web seve module that makes use of a set of dynamc lbaes, known as n-coe API (Applcaton Pogammng Inteface). The API offes the functonaltes of the ZgBee netwok. The web seve module mplements the nnovatve locatng technque pesented n ths pape. Ths way, n-coe Polas ncopoates the descbed Mult-Laye Peceptons n ode to mtgate the gound eflecton effect. On the one hand, the compute gathes the detecton nfomaton sent by the n-coe Sus D actng as eades to the coodnato node. One the othe hand, the compute acts as a web seve offeng the locaton nfo to a wde ange of possble clent ntefaces. In addton, the web seve Fg. 4. Basc achtectue of the n-coe Polas RTLS. The opeaton of the system s as follows. Each use o object to be located n the system caes an n-coe Sus- B actng as tag. Each of these tags boadcasts peodcally a data fame ncludng, among othe nfomaton, ts unque dentfe n the system. The est of the tme these devces ae n a sleep mode, so that the powe consumpton s educed. Ths way, battey lfetme can each even seveal yeas, egadng the paametes of the system (boadcast peod and tansmsson powe). A set of n-coe Sus D devces s used as eades thoughout the envonment, beng placed on the celng and the walls. The boadcast fames sent by each tag ae eceved by the eades that ae close to them. Ths way, eades stoe n the memoy a table wth an enty pe each detected tag. Each enty contans the dentfe of the tag, as well as the RSSI (Receved Sgnal Stength Indcaton) and the LQI (Lnk Qualty Indcato) gatheed fom the boadcast fame ecepton. Peodcally, each eade sends ths table to the coodnato node connected to the compute. The coodnato fowads each table eceved fom each eade to the compute though the USB pot. Theefoe, usng all these detecton nfomaton tables, the n-coe API apples a set of locatng technques, ncludng that poposed n ths pape, n ode to estmate the poston of each tag n the montoed envonment. Then, the web seve module offes the locaton data to emote clent ntefaces as web sevces though HTTP (Hypetext Tansfe Potocol) ove SOAP (Smple Object Access Potocol). Ths way, the n-coe Polas system ncludes thee basc clent ntefaces: a desktop applcaton, a web applcaton and a moble applcaton. Fgue 5 shows a sceenshot of the web clent nteface. All these clent ntefaces has been desgned to be smple, 761

7 ntutve and easy-to-use. Though the dffeent ntefaces, admnstato uses can watch the poston of all uses and objects n the system n eal-tme. Futhemoe, admnstatos can defne estcted aeas accodng to the uses' pemssons. Ths way, f some use entes n an aea that s fobdden to t egadng ts pemssons, the system wll geneate an alet that s shown to the admnstato o monto use though the clent ntefaces. In addton, such alets ae egsteed nto the database, so admnstatos can check anytme f any use volated ts pemssons. Lkewse, admnstatos can quey the database n ode to obtan the locaton tack of a cetan use, obtanng statstcal measuements about ts moblty o the most fequent aeas whee t moves. Fg. 5. Web clent nteface of the n-coe Polas RTLS. Futhemoe, uses can use one of the geneal-pupose buttons povded by the n-coe Sus B devces n ode to send an alet to the system. Smlaly, admnstatos can send alets fom the system to a use o a set of uses, whch can confm the ecepton usng othe of the buttons. The system not only povdes locatng featues, but also schedulng and automaton functonaltes. The system can be ntegated wth senso and actuato devces. These devces, called n-coe Sus A, ncopoate seveal communcaton pots (GPIO, ADC, I2C and UART though USB o DB-9 RS-232) to be connected to dstnct devces, ncludng almost evey knd of senso and actuato [22]. By means of the automaton engne povded by the n-coe API, the n-coe Polas system can schedule automaton tasks, as well as monto all sensos n the envonment n eal-tme. All the nfomaton can be accessed though the dstnct web clent ntefaces. 5 Conclusons and futue wok Among the wde ange of Weless Senso Netwoks applcatons, Real-Tme Locatng Systems ae emegng as one of the most exctng eseach aeas. Healthcae, suvellance o wok safety applcatons ae only some examples of the possble envonments whee RTLSs can be exploted. Thee ae also dffeent weless technologes that can be used on these systems. The ZgBee standad offes nteestng featues ove the est of technologes, as t allows the use of lage mesh netwoks of low-powe devces and the ntegaton wth many othe applcatons as t s an ntenatonal standad wokng on unlcensed fequency bands. The opeaton of Real-Tme Locatng Systems can be affected by undesed phenomena such as the multpath effect, and moe specfcally, the gound eflecton effect. Ths pape poposes a new mathematcal model amed at mpovng the pecson of RTLSs based on WSNs. The use of Atfcal Neual Netwoks to foecast dstances fom RSSI levels allows mpovng the estmaton of dstances. In addton, focusng the foecast accodng to tme sees allows educng the gound eflecton effect that occus when consdeng only the last RSSI measuement. The use of measuements fom seveal eades as nputs of the MLP n the poposed model also educes the pedcton eo. Ths way, the gound eflecton effect s mtgated and theefoe the appoxmatons povded by othe methods wth hgh adjustment goodness, as the logathmc egesson model wee mpoved. Ths mpovement n the dstances foecastng s vey elevant to estmate the postons of the tags, thus optmzng the oveall calculatons of locatng technques. As futue wok t s planned the educton of the numbe of eades necessay to pefom the locatng pocess, as well as the mplementaton n lage envonments. Theefoe, the system wll subjected to an ntensve set of expements n ode to valdate both the mathematcal model and the system. Detals of these expements and the esults wll be publshed n futhe complementay papes. Futue wok also ncludes the study of moe detaled multpath models such as Rcean and Raylegh fadng o shadowng [23]. Acknowledgments. Ths poject has been suppoted by the Spansh Mnsty of Scence and Technology poject TIN C03-03: Oganzacones Vtuales Adaptatvas: Mecansmos, Aqutectuas y Heamentas (OVAMAH). We also want to thank Nebusens, S.L. fo ts suppot n developng and testng the n-coe Polas RTLS. Refeences [1] J. K. Ray, M. E. Cannon, and P. C. Fenton, Mtgaton of statc cae-phase multpath effects usng multple closely spaced antennas, Navgaton- Washngton, vol. 46, no. 3, pp , [2] C. Neguzan, C. Despns, and S. Affès, Indoo Geolocaton wth Receved Sgnal Stength Fngepntng Technque and Neual Netwoks, n Telecommuncatons and Netwokng - ICT 2004, vol. 3124, Spnge Beln / Hedelbeg, 2004, pp [3] D. I. Tapa, J. F. de Paz, S. Rodguez, J. Bajo, and Juan M. Cochado, Mult-Agent System fo Secuty Contol on Industal Envonments, Intenatonal Tansactons on System Scence and Applcatons, vol. 4, no. 3, pp ,

8 [4] Hu Lu, H. Daab, P. Banejee, and Jng Lu, Suvey of Weless Indoo Postonng Technques and Systems, Systems, Man, and Cybenetcs, Pat C: Applcatons and Revews, IEEE Tansactons on, vol. 37, no. 6, pp , [5] K. Kaemaungs and P. Kshnamuthy, Modelng of ndoo postonng systems based on locaton fngepntng, n INFOCOM Twenty-thd Annual Jont Confeence of the IEEE Compute and Communcatons Socetes, vol. 2, pp vol.2, [6] A. Goldsmth, Weless communcatons. Cambdge Unvesty Pess, [7] L. W. Baclay and I. O. E. Engnees, Popagaton of adowaves. IET, [8] Eu Seok Km, Jn Ik Km, Ik-Seon Kang, Chan Gook Pak, and Jang Gyu Lee, Smulaton esults of angng pefomance n two-ay multpath model, n Contol, Automaton and Systems, ICCAS Intenatonal Confeence on, pp , [9] Bn Dng, L Chen, Danlong Chen, and Hatao Yuan, Applcaton of RTLS n Waehouse Management Based on RFID and W-F, n Weless Communcatons, Netwokng and Moble Computng, WCOM '08. 4th Intenatonal Confeence on, pp. 1-5, [10] M. A. Stelos, A. D. Nck, M. T. Effe, K. M. Dmts, and S. C. A. Thomopoulos, An ndoo localzaton platfom fo ambent asssted lvng usng UWB, n Poceedngs of the 6th Intenatonal Confeence on Advances n Moble Computng and Multmeda, pp , [11] G. Galat, M. Gasbaa, P. Magao, P. De Maco, L. Mene, and M. Pc, New Appoaches to Multlateaton pocessng: analyss and feld evaluaton, n Rada Confeence, EuRAD d Euopean, pp , [15] V. N. Vapnk, Statstcal Leanng Theoy. Wley- Intescence, [16] A. J. Smola and B. Schölkopf, A tutoal on suppot vecto egesson, Statstcs and Computng, vol. 14, no. 3, pp , [17] S. A. Kalogou, Applcatons of atfcal neual netwoks n enegy systems: A evew, Enegy Conveson and Management, vol. 40, no. 10, pp , Jul [18] D. E. Rumelhat, G. E. Hnton, and R. J. Wllams, Leanng epesentatons by back-popagatng eos, Natue, vol. 323, no. 6088, pp , Oct [19] A. Mata and J. M. Cochado, Foecastng the pobablty of fndng ol slcks usng a CBR system, Expet Systems wth Applcatons, vol. 36, no. 4, pp , [20] P. Anand, B. V. N. Sva Pasad, and C. Venkateswalu, Modelng and optmzaton of a phamaceutcal fomulaton system usng adal bass functon netwok, Intenatonal Jounal of Neual Systems, vol. 19, no. 2, pp , Ab [21] H. Katsuua and D. Speche, Computatonal aspects of Kolmogoov's supeposton theoem, Neual Netwoks, vol. 7, no. 3, pp , [22] n-coe: A Faste and Ease Way to Ceate Weless Senso Netwoks. Reteved May 8, 2011, fom [23] A. Schmtz and M. Weng, The effect of the ado wave popagaton model n moble ad hoc netwoks, n Poceedngs of the 9th ACM ntenatonal symposum on Modelng analyss and smulaton of weless and moble systems, pp , [12] J. Xe, R. Palme, and D. Wld, Multpath mtgaton technque n RF angng, n Electcal and Compute Engneeng, Canadan Confeence on, pp , [13] Z. Salcc and E. Chan, Moble Staton Postonng Usng GSM Cellula Phone and Atfcal Neual Netwoks, Weless Pesonal Communcatons, vol. 14, no. 3, pp , Sep [14] Y. Lecun, L. Bottou, G. O, and K. Mülle, Effcent BackPop, n Lectue Notes n Compute Scence, vol. 1524, Spnge Velag, 1998, pp