Indoor Location based on Particle Swarm Optimization and BP Neural Network

Size: px

Start display at page:

Download "Indoor Location based on Particle Swarm Optimization and BP Neural Network"

Gwen Robinson
5 years ago
Views:

1 Inoor Locaton base on Partcle Swarm Optmzaton an BP Neural Network Abstract Shanshan Chen a, *, Zhca Sh b Department of school of Electrc an Electronc Engneerng, Shangha Unversty of Engneerng Scence, Shangha, Chna a chenshanshan3310@163.com, b @qq.com The eal postonng methos are of great mportance n pervasve computng servces. RFID technology wth low cost an smple structure s often use for noor postonng. Base on the avantages of RFID technology, an noor postonng metho base on partcle swarm to optmze BP neural network an reaer eployment s propose. Snce precton results of tratonal BP neural network are easy to fall nto local optmum, ths metho uses PSO to optmze the weghts an threshols of BP neural networks to avo the shortcomngs. The partcle swarm optmzaton s use to optmze the reaer's poston, so that the Euclean stance between the reference tags s approprate an the tolerance of the postonng system to sgnal fluctuatons can be mprove. In aton, n orer to mprove the postonng accuracy, the expermental ata are processe by Gauss flterng metho. The expermental results show that the propose metho has better performance than tratonal methos. Keywors Rao frequency entfcaton; partcle swarm optmzaton algorthm; BP neural network; reaer eployment; noor poston. 1. Introucton Wth the rap evelopment of wreless sensor, noor postonng nformaton servce has become a hot topc. Global Postonng System(GPS) s ffcult to meet the requrement of precson postonng uner the noor envronment because t lacks of lne of sght between the satellte recever an satellte transmsson [1]. In recent years, researchers have put forwar a varety of postonng technology, such as nfrare, ultrasonc, rao frequency entfcaton (RFID), UWB, WIFI, Bluetooth an ZgBee. Among them, the RFID technology s sutable because ts prce s very low an ts ata transmsson s base on rao-wave communcatons. The concept of RFID, whch s recently use n numerous nustral applcatons from asset trackng to supply chan management [2 6], has receve sgnfcant attenton among the researchers [7]. The tratonal RFID postonng metho can be manly ve nto two sorts: geometrc methos an the algorthms base on RSSI [8]. A seres of researches relate to geometrc methos have been conucte recently. Shen et al. use TOA measurements n noor envronments to estmate the locaton of passve object [9]. The TOA metho can acheve hgh accuracy of rangng n the lne-of-sght an multpath envronment, but t nees accuracy of hgh clock synchronzaton whch s expensve between the transmttng en an recevng en. Jung et al. employe TDOA localzaton algorthm n an evaluate the performance of the propose localzaton metho [10]. Wen an Lang propose an noor AOA estmaton algorthm, the auto-focusng metho s frst use to obtan a coherently combne matrx among fferent subcarrer frequences an then the stuy mplement Toepltz processng n the coherently combne matrx to present the receve multpath sgns n spatal oman. The propose metho outperforms the conventonal auto-focusng metho wth low error bas, even though all multpath sgnals are hghly correlate [11]. 283

The geometrc metho s to obtan the stance between the tag an the reaer base on the relatonshp between sgnal strength loss an stance n the propagaton moel [12].

The metho s senstve to ntal weght an threshol senstvty, easy to fall nto local optmal an slow convergence spee.

Therefore, we propose to optmze the parameters an reaer eployment of BP neural network algorthm wth partcle swarm algorthm to mprove the postonng effect.

The remanng part of ths paper s organze as follows: In Secton 2 we state our approach an present our algorthm. In Secton 3 we optmze the poston of reaer.

2 The geometrc metho s to obtan the stance between the tag an the reaer base on the relatonshp between sgnal strength loss an stance n the propagaton moel [12]. But noor sgnal transmsson s unstable, the results base the sgnal propagaton moel are lack of applcablty n the noor envronment. At present, the BP neural network s usually use n noor postonng. The metho s senstve to ntal weght an threshol senstvty, easy to fall nto local optmal an slow convergence spee. In aton, the exstng metho oes not conser the eployment of RFID reaer when the algorthm s mprove. Therefore, we propose to optmze the parameters an reaer eployment of BP neural network algorthm wth partcle swarm algorthm to mprove the postonng effect. The expermental results show that the postonng accuracy an stablty of BP neural network algorthm are better than orgnal postonng. The remanng part of ths paper s organze as follows: In Secton 2 we state our approach an present our algorthm. In Secton 3 we optmze the poston of reaer. Secton 4 emonstrates the harware expermental metho an the results. Fnally, the conclung remarks are mae n Secton Inoor Postonng System 2.1 RFID system RFID noor postonng systems nclue reaers, reference tags, antennas an servers [13], as shown n Fg.1. The RFID reaer s responsble for powerng an communcatng wth a tag. RFID tags are wely apple n many nustres, for example, an RFID tag attache to an automoble urng proucton can be utlze to montore ts progress n the assemblng, RFID-tagge contaners can be tracke urng the transportaton [14,15]. Unlke actve RFID tags that are powere by batteres, passve RFID systems however communcate through the backscatter rao lnks ue to that passve tags (no batteres powere) can only passvely collect energy from the n-ar backscattere rao sgnal. So we select passve RFID tag n our experment. The RFID antenna captures energy an transfers the tag s ID (the tag s chp coornates ths process). The RFID mleware can process the receve ata an calculate the locaton nformaton. RFID tag RFID antenna RFID mleware RFID reaer Fg. 1 RFID postonng system. 2.2 RSSI varaton ue to multpath shaowng The frequency of RFID s generally n the range of 125kHz to 5.8GHz. Ths paper uses passve RFID Ultra Hgh Frequency (UHF) tags whch on t requre the battery to prove energy an reach about 20 meters of communcaton range n the open conton. The tags shoul be actvate by the nterrogaton an reply by senng a unque entfcaton strng back at the reaer [16]. Then the reaer wll receve the ata of the tag feeback, nclung tme, RSSI value an ID nformaton. The sgnal strength of the reference tag s not a stable value. The sgnal strength of the same reference tag wll be affecte by a lot of noor nterference factors, an the fluctuatons cannot be gnore, because the fngerprnt must truly reflect the sgnal strength characterstcs an can guarantee the user's locaton estmaton accuracy n the onlne phase [17]. RSSI of the reaer for acqurng the postonng tag ecreases wth the ncrease of the propagaton stance, an the relatonshp between the RSSI value an the propagaton stance s shown n Eq. (1). RSSI ( ) ( ) M j( m ) m e (1) m 1 m

3 Where s the stance between the reaer an tag, s the path coeffcent, M s the total number of reflectons, s the coeffcent of the m th, an s the length of the m th reflecton path. m In the actual scene, the noor obstacle wll affect the postonng result, whch leas to the phenomena of sgnal reflecton, ffracton an scatterng. The common sgnal path loss moel n noor envronment s the logarthmc moel represente by Eq. (2). PL( ) PL( 0)+10 log10 X (2) Where s path loss for reference stance, s path loss exponent an enotes a Gaussan ranom nose wth zero mean an stanar evaton of. For a gven physcal envronment, the precte loss path moel can be analyze from the measure ata. As shown n Fg.2, there s some screpancy between the actually measure value of RSSI an the value obtane from the loss path moel. PL( 0) m 0 0 X Fg.2 Relaton between RSSI an stance. 2.3 BP neural network optmze by Partcle Swarm Optmzaton algorthm In orer to mprove the performance of BP neural network, we use Partcle Swarm Optmzaton algorthm (PSO) to fn the weghts an threshols of the BP neural network. PSO s a typcal representatve of swarm ntellgent algorthm to search for the best soluton by smulatng the movement of flockng of brs. Each partcle n the PSO algorthm represents the weght an threshol of the BP network, an each partcle correspons to a ftness value etermne by the ftness functon. [18]The sum of the absolute value of the precton error of tranng ata s taken as the nvual ftness value. The smaller the nvual ftness value s, the better the nvual s. The PSO algorthm obtans the optmal weghts an threshols of the network by fnng the mnmum ftness value, where the ftness functon s shown n Eq. (3) [19]. N M 1 2 Ft ( Op, q op, q ) (3) N p 1 q 1 Where Ft s the partcle ftness functon; N s the number of tranng samples, M s the number of output noes of the neural network, O pq, s the expecte output value of the q th noe of the th sample, o pq, s the actual output of the q th noe of the p th sample value. Partcles keep upatng spee an poston whle lookng for space untl the ftness functon value reaches the set ftness value. Partcle velocty an poston are contnuously upate accorng to Eq. (4) an Eq. (5). 285 p

4 Where u socal coeffcent, partcle at the p g v ( t 1 ) uv ( t) c1ran 1( p x ( t) c2ran 2( pg x ( t)) (4) x ( t 1) x ( t) v ( t 1) (5) estmate the nerta of the partcle parameters; t th; ran1 an x c1 an ran2 are ranom values between 0 an 1. s the poston of the partcle at the t th; p c2 enote cogntve coeffcent an v s the velocty of the s best locaton for all the partcles; s the best locaton. Postonng process s vsble nto offlne phase an onlne phase. In the offlne phase, we establsh the postonng moel of noor poston. In the onlne phase, we collect RSSI values after Gaussan flterng to obtan the GRSSI vector. The GRSSI vector s taken as the nput of the postonng moel an algorthm s the estmate poston of the tag. 3. PSO algorthm for reaers eployment ( xy, ) obtane by the postonng Reaer eployment s usually base on experence whch oes not get best performance of poston. In orer to mprove the accuracy of poston, we use PSO to optmze the reaer eployment. 3.1 Deploy reaers When the number of reaers s etermne, reaers are eploye accorng to the prncple that sgnal space Euclean stance shoul be as bg as possble whle varance of sgnal space Euclean stance as small as possble. The value of RSSI measure at same pont urng postonng fluctuates of Gaussan strbuton constantly. Therefore, the postonng result s relate to the Euclean stance of the sgnal space between reference tags. The Euclean stance of a reference tag refers to the average Euclean stance of all reference labels wthn 2 meters of the reference pont, where the Euclean stance of the sgnal space s efne as follows: 1 2 P 1 2 P If RSSI s ( RSSI s, RSSI s,, RSSI s ) an RSSI t ( RSSI t, RSSI t,, RSSI t ) are the Pth reaer receve the RSSI vector of the Sth an the Tth reference tags respectvely, the Euclean stance of the sgnal spaces between the Sth an the Tth reference tags s gven by Eq. (6) P k k S, T S T k 1 2 D ( RSSI RSSI ) (6) Due to the real-tme varaton of postonng envronment, the sgnal receve at a certan locaton s a varable value rather than a efnte value. Suppose a sgnal receve by a reference tag s a ranom varable, whch changes n a crcle centere at a pont O an n a raus r, as shown n Fg.3. A O r B -1Bm -2Bm -3Bm -4Bm -5Bm -6Bm 2m A O r B -1Bm -3Bm -5Bm -7Bm -9Bm 2m -11Bm Fg.3 The nfluence of sgnal Euclean stance on postonng accuracy. 286

5 As shown n Fg. 4, when the physcal stances between reference labels A an B are both 2 meters, the larger the European stance s, the smaller the postonng error s. Therefore, the tolerance of the postonng system to sgnal fluctuatons becomes stronger. In orer to acheve the above goal, the optmzaton of reaer eployment nees to make the fference between the average Euclean stance of all reference tags an ts stanar evaton be the maxmum, an the fference between two tags s calculate as shown n Eq. (7). L DST, T F 2 max H h sqrt( (( ) h) ) (7) length( F) S 1 Where H s the sgnal space Euclean stance varance of all the reference tags n the target areas, s the sgnal space Euclean stance average of all reference tags, F s a pont set n whch the ponts are less than meters from the th reference tag, D s k-mensonal sgnal space Euclean stance between the th reference tags an th reference tags, L s the number of all reference tags. h j 3.2 Reaer Deployment Optmzaton The essence of reaer eployment s to make use of the sze of the European stance of the sgnal space between reference tags to etermne whether the sgnal coverage characterstcs are favorable to the postonng nees of the system. In orer to obtan the optmal soluton of reaer eployment, we use partcle swarm optmzaton algorthm to get the optmal soluton of reaer eployment. In orer to PSO algorthm s use to optmze the eployment of the reaer, each partcle n the PSO algorthm represents the poston of the reaer, an each group of partcles correspons to a ftness value etermne by the ftness functon. The fference between the mean Euclean stance of all reference tags an ts stanar evaton s taken as the ftness value of an nvual, an the better the nvual ftness value s, the better the nvual s. The PSO algorthm obtans the optmal poston of the reaer by fnng the maxmum ftness value, whch s use as the partcle swarm ftness value n Eq. (7). Partcles keep upatng spee an poston whle lookng for space untl the ftness functon value reaches the set ftness value. Partcle velocty an poston are contnuously upate accorng to Eq. (4) an Eq. (5). 4. Smulaton results an analyss In orer to verfy the feasblty of PSO to the eploy reaers s optmze an test the stablty of optmzng BP neural network algorthm optmze by PSO, we smulate experments by MATLAB 2016a. The followng Fg.4 shows the area of 16m * 16m wth 256 reference labels an four reaers (the stance of ajacent reference tag s 1meters). reference tag reaer Fg.4 Intal eployment of the reaer. 287

6 4.1 Smulaton scene The wreless sgnal propagaton emprcal formula s use to smulate the RSSI value between the reference tag an the reaer. In Eq. (8), enotes the sgnal strength of the reference tag when t s from the reaer meter an E ncates the transmtte energy. In the experment, the esre parameters are set as the E 0Bm, Bm an an we a Gaussan nose wth zero mean an stanar evaton of 3 use to smulate the actual envronment ue to sgnal reflecton of walkng an other factors on the RSSI value. RSSI ( ) E PL( 0) 10 log 10( ) X (8) PL( ) RSSI ( ) 0 In the process of smulaton, the most commonly use evaluaton stanar s poston error calculate by the Eq. (9), where th tracke tag an poston of the th tracke tag. 2 2 error = ( x x ) ( y y ) (9) ( x, y)s the actual poston of the o o 0 j ( x, y )s the precte 4.2 Reaer eployment performance We use the PSO algorthm to optmze the eployment poston of the reaer. Fg.4 shows the ntal eployment of the reaer. Fg.5 shows the eployment of the reaer after optmzaton. o o reference tag reaer Fg.5 Optmze the eployment of the reaer. After optmzng the poston of the reaer through the PSO, the sgnal strength vectors of the reference tags an ther corresponng coornate values are use as the two nputs respectvely as the tranng samples of the BP neural network an the PSO-BP algorthm. In aton, compare wth the prevous expermental results, we can get four fferent postonng methos as follows: Case1: BP algorthm postonng metho wthout reaer optmzaton Case2: PSO-BP algorthm postonng metho wthout reaer optmzaton Case3: BP algorthm postonng metho of optmzng reaer eployment Case4: PSO-BP algorthm postonng metho of optmzng reaer eployment The sgnal strength vector of the tracke tag s taken as the test nput ata of the BP neural network an the PSO-BP. After the moel operaton has been trane, the poston of the tracke tag can be obtane. Fg.6 s the frst 50 tmes the test ata error comparson chart to show the preference of four fferent postonng methos. Fg.7 s the four postonng metho of postonng error cumulatve strbuton functon. It can be seen from Fg.10 shows that the preference of poston s relate to choce of algorthm an eployment of the reaer. Case4: PSO-BP algorthm postonng metho of optmzng reaer eployment has the least overall error. Fg.7 shows the cumulatve strbuton functon n four fferent cases. The abscssa ncates the postonng error, an the ornate ncates the cumulatve probablty. As shown n Fg.7, the Case4 metho has better postonng effect. 288

7 Fg.6 Comparson of four postonng methos. Fg.7 The cumulatve strbuton functon of the four methos. Table 1. Comparson of four kns of postonng methos. Case1 Case2 Case3 Case4 Probablty=30% Probablty=60% Probablty =85% Mean error(m) St Table 1. compares the four fferent postonng algorthms. As seen from Table 1, the error of PSO-BP algorthm postonng metho of optmzng reaer eployment s smaller than that of the other three methos at the same cumulatve probablty of 85%. The average error of BP algorthm postonng metho of optmzng reaer eployment s m less than BP algorthm postonng metho wthout reaer optmzaton an the average error of the PSO-BP of optmzng reaer eployment s m less than the average error of the PSO-BP wthout reaer optmzaton. Therefore, the preference of reaer eployment wth PSO s better than reaer eployment wthout algorthm optmzaton. In aton, through the comparson of Table 1, t can be seen that usng PSO to optmze BP neural network has hgher postonng accuracy an better postonng stablty than 289

8 BP neural network. Therefore, the fourth metho has better postonng accuracy an stablty than the other three methos. 5. Concluson In the RFID postonng, the BP neural network s rectly use for postonng, an the postonng accuracy an stablty of the RFID postonng nees to be mprove. In aton, the postonng effect of the reaer eployment base on experence s also poor. Therefore, ths paper proposes to use PSO algorthm to optmze the parameters of BP neural network an reaer's poston respectvely. On the one han, by optmzng weghts an threshols of BP neural network, the moel precton results are avoe from gettng nto local optmum. On the other han, the partcle swarm optmzaton algorthm optmzes the eployment of reaers, whch mproves the tolerance of sgnal postonng system. From the expermental results we can see that PSO algorthm to optmze the parameters of BP neural network an reaer's poston respectvely shows an average postonng error m less than the average error of BP algorthm postonng metho wthout reaer optmzaton. Ths paper mproves the postonng accuracy an stablty from these two aspects. So the partcle swarm optmzaton algorthm whch optmzes BP neural network parameters an reaer eployment s more sutable for noor postonng. Acknowlegments Ths stuy was supporte n part by Natonal Scence Fun for Young Scholars No , by Innovaton Project of Shangha Unversty of Engneerng Scence No.17KY0202. Authors contrbutons: The stuy of the moble multmea crow servce cooperaton control protocol was carre out by Shanshan Chen, an the revson of wavelet moel was one by Zhca Sh an Fe Wu. The smulaton experment an cong work were one by all the authors. Ths manuscrpt ha been prepare an checke by both of the authors together. All authors rea an approve the fnal manuscrpt. Conflcts of Interest: The authors eclare no conflct of nterest. References [1] A. Montaser, O. Moselh, RFID noor locaton entfcaton for constructon projects. Automat. Constr. 39 (2014) [2] Y.J. Zuo, Survvable RFID systems: Issues, challenges, an technques. IEEE Trans. Syst., Man, Cybern. C. 40 (2010) [3] Ganno, F. B. Montruccho, M. Rebauengo, Sanchez, E.R. On mprovng automaton by ntegratng RFID n the traceablty management of the agr-foo sector. IEEE Trans. In. Electron. 56 (2009) [4] T.M. Cho, Coornaton an rsk analyss of VMI supply chans wth RFID technology. IEEE Trans. In. Informat. 7 (2011) [5] J.D. Porter, D.S. Km, An RFID-enable roa prcng system for transportaton. IEEE Syst. J. 2 ( 2008) [6] H.H. B; D.K. Ln, RFID-enable scovery of supply networks. IEEE Trans. Eng. Manag. 56 (2009) [7] Y. Son, M. H. Joung, Y.W. Lee, O.H. Kwon, H.J. Song, Tag localzaton n a two-mensonal RFID tag matrx. Future Gener. Comp. Sy [8] C. Fguera, J.L. Rojo-Álvarez, M. Wlby, I. Mora-Jménez, A.J. Caamaño, Avance support vector machnes for noor locaton. Sgnal Process. 92 (2012) [9] J. Shen, A.F. Molsch, J. Salm, Accurate Passve Locaton Estmaton Usng TOA Measurements. IEEE Trans. Wrel. Commun. 11 (2011) [10] S.Y. Jung, S. Hann, C.S. Park, TDOA-base optcal wreless noor localzaton usng LED celng lamps. IEEE Trans. Consum. Electr. 57 (2012)

9 [11] F. Wen, C. Lang, An Inoor AOA Estmaton Algorthm for IEEE ac W-F Sgnal Usng Sngle Access Pont. IEEE Commun. Lett. 18 (2014) [12] R.C. Chen, S.W. Huang, Y.C. Ln, Q.F. Zhao, An noor locaton system base on neural network an genetc algorthm. Int. J Sens. Netw. 19 ( 2015) [13] Z. Zhang, Lu, Z. V. Saakan, X. Qn, Q. Chen, L. Zheng, Item-level noor localzaton wth passve uhf rf base on tag nteracton analyss. IEEE Trans. In. Electron. 61 (2013) [14] F. Rzzo, M. Barbon, L. Faggon, G. Azzaln, M. Sron, Improve securty for commercal contaner transports usng an nnovatve actve rf system. J. Netw. Comput. 34(2011) [15] L.A. Amaral, F.P. Hessel, E.A. Bezerra, J.C. Corrêa, O.B. Longh, T.F. Das, eclourf a moble software framework archtecture for p.ervasve rf-base applcatons. J. Netw. Comput. 34 (2011) [16] S.S. Saab, Z.S. Naka, A Stanalone RFID Inoor Postonng System Usng Passve Tags. IEEE Trans. In. Electron. 58 (2011) [17] C. Ren, N. An, J. Wang, L. L, B. Hu, D. Shang, Optmal parameters selecton for BP neural network base on partcle swarm optmzaton: A case stuy of wn spee forecastng. Knowl.-Base Syst. 56 (2014) [18] A.M. Pnto, A.P. Morera, P.G. Costa, A Localzaton Metho Base on Map-Matchng an Partcle Swarm Optmzaton. J Intell. Robot Syst. 77 (2015) [19] Y. He, W. Meng, L. Ma, Z. Deng, Rap eployment of APs n WLAN noor postonng system. Internatonal ICST Conference on Communcatons an Networkng, Harbn, Chna, 2011,pp

Improvements of Indoor Fingerprint Location Algorithm based on RSS

Improvements of Indoor Fingerprint Location Algorithm based on RSS Internatonal Journal of Scence Vol.4 No.1 017 ISSN: 1813-4890 Improvements of Indoor Fngerprnt Locaton Algorthm based on RSS Quyue Zhu a, Qang Yu b, Q Lu c and Kun Sh d School of Computer and Software