www.ijcsi.org http://dx.doi.org/10.20943/01201604.1115 11 Data Miig of Bayesia Networks to Select Fusio Nodes from Wireless Networks Yee Mig Che 1 Chi-Shu Hsueh 2 Chu-Kai Wag 3 1,3 Departmet of Idustrial Egieerig ad Maagemet, Yua Ze Uiversity 135 Yua-Tug Rd., Chug-Li, Tao-Yua, Taiwa, R.O.C. 2 Iformatio ad Commuicatio Research Divisio, Natioal Chug Sha Istitute ad Techology, Tao-Yua, Taiwa, R.O.C. Abstract I the wireless sesor etworks(wsn) maagemet, uder differet regios ad differet time, spectral data is very big, resultig i the sesor odes maagemet would be difficult, this paper based o Bayesia etwork of data miig to select fusio ode i the wireless sesor etworks, which focused o balacig eergy cosumptio. Uder the architecture of Bayesia etwork, i which ca itegrate the quatitative ad qualitative kowledge ito a comprehesive probabilistic kowledge represetatio ad iferece WSN eviromet. It discusses how these techiques ca prepare miig the wireless sesor etworks data iside the etwork (i-etwork) before data fusio further processig as big data. Keywords: Data miig, Bayesia Networks, Data fusio 1. Itroductio Wireless sesor etworks (WSN) cosist of a collectio of a large umber of small, distributed over a large area, ad low powered sesor odes capable of commuicatig with each other via a ad-hoc wireless etwork. Each WSN cosists of three primary compoets: sesor odes, fusio ode, ad a remotely base statio. The sesor odes are resposible for collectig the locally available sesor data. The sesor odes are small ad iexpesive. Sice most odes are traditioally battery powered, power cosumptio is a importat cosideratio whe settig up a WSN. Oce the data has bee collected from the sesor odes, they the trasmit that iformatio to a fusio ode [1]. The brai of a WSN is usually a decisiomakig algorithm that is capable of correctly mappig a set of ewly collected observatios from the sesors to oe or more predefied fusio ode [2]. Due to their limited power ad short commuicatio rage, the iformatio that the odes sed to the remotely base statio is usually put through a data fusio techique before beig set to the base statio. This allows for the data to be more accurate ad also reduces overhead i the etwork[2]. The eed for extractig kowledge from the sesor data, collected from WSN, has also become a importat issue i real-time decisio algorithms. I additio, the rapid chage of the moitored data requires the implemetatio of data miig algorithms i order to get a reasoable time respose or predictio. A imbalaced eergy problem may have aother implicatio i WSN where it could be a symptom of producig traffic hot-spot i WSN. The eergy cosumptio i the sesors may become imbalaced too, which leads to drai out for some local odes. Some data miig solutio has bee selected to better fusio odes ad traffics although it is aimed at the eergy level. The data miig algorithms could be geerally classified ito either a cetralized or a distributed data processig [3]. As a distributed data processig model for the probabilistic relatioships amog a set of variables, the Bayesia etwork (BN) has become a popular represetatio for ecodig ucertai expert kowledge i data miig domais over the last decade [4]. This paper use data miig approach, Bayesia Networks, to select a fusio ode. The outlie of this paper is as follows. Distributed sesors data miig processig approaches are preseted i Sectio 2. Developed fusio ode selectio usig Bayesia etworks are stated i Sectio 3. Simulatio results o data fusio discover fusio ode are preseted i Sectio 4. Fially, coclusios are made i Sectio 5. 2. Distributed s Data Miig Processig Approaches I the distributed sesors data processig approach, each ode uses its limited computig resources to perform the miig process. The process of acquirig the data is
www.ijcsi.org http://dx.doi.org/10.20943/01201604.1115 12 aother factor impactig the data quality, because the trasmissio of the sigal is affected by all kids of ucertai factors. Thus, ucertaity is a iheret property of the WSN data. The Bayesia method itegrates a prior kowledge about the targets uder study ad the iformatio provided by ew data set, followed by ecodig the multi-kowledge ito coditioal probability etwork model. Thus, Bayesia etwork i cojuctio with Bayesia statistical techiques facilitates the combiatio of domai kowledge with the relevat data. The sytax of Bayesia etwork is clearer, which ca reaso i dual directio ad ca be costructed ad debugged rapidly. The disadvatage of Bayesia etwork is that the computatio complexity is high. Bayesia Network Sytax illustrated as the followig [5]: BN= (Structure, CPT) (1) Structure cotai odes ad arcs Nodes: radom variable. (a) Nodes ca be cotiuous or discrete. (b) Nodes ca have two states or more. (c) Nodes ca be determiistic or odetermiistic. Arcs: relatioships betwee odes. (a) Arcs represet causal relatioships of odes. (b) Arc betwee x ad y represets that x has direct causal ifluece oly. (2) CPT: Coditio Probability Table (a) Each ode has coditio probability which is stored i a table (CPT). (b) Value i table is P(Xi parets(xi)), parets(xi) is the set of paret odes of Xi. (c) Root ode is particular, as it has o paret ode ad has oly prior probability: parets (X i ) = Φ, so P(X i parets(x i ))= P(X i ). Suppose we have two variables E ad H. If H has bee realized (i.e. we kow exact value of H), we might wat to kow what is probability of the evet E. The situatio where we are dealig with probability of oe evet, give that aother evet has occurred, is called coditioal probability. Mathematically coditioal probability is defied: P(E H k ) = P(E H k) P(H k ). (1) If E 1, E 2,..., E are mutually exclusive evets such that, i=1 E i = Ω, E i s are said to be exhaustive. Two variables are said to be disjoit if they have o elemets i commo. If variables are disjoit ad exhaustive Equatio (2) holds: E = (E H i ) i (E H i ) (E H j ) =, i j (2) The theory of the Bayesia etworks assumes that evets are disjoit ad exhaustive. If they are ot, the results are ot cosistet. Whe evets are disjoit ad exhaustive, the probability of E ca be calculated via coditioal probabilities: P(E) = i=1 P(E H i ) = i=1 P(E H i ) P(H i ) (3) Usig Equatio (1) we ca express the sum of the itersectio of E ad H as follows. P(E H k ) = P(E H k ) P(H k ) = P(H k E) P(E) (4) Now it is possible to place Equatio.(3) to Equatio.(4) ad obtai Equatio (5). P(H k E) = P(E H k) P(H k ) P(E) = P(E H k) P(H k ) i P(E H i ) P(H i ) Equatio (5) is widely kow as Bayes formula. Bayes formula is the foudatio of Bayesia etworks that are actually othig but a etwork structure where observatios are hadled usig Bayes theorem. I the Equatio (5), H k meas all hypotheses (subscript k refers to the fact that there are several hypotheses). Probabilities are a priori values from experts. A term P(H k ) is called a priori probability. The deomiator ca be cosidered as a ormalizig factor which ormalizes the probability betwee zero ad oe. 3. Fusio odes selectio i WSN usig BN For WSN, these sesor odes ot oly detect the target, but they also collect the data process ad trasmit it to the outside world for further processig. These sesor odes require careful resource maagemet as they are tightly costraied i terms of power, trasmissio power, processig capacity ad storage capability. The data beig sesed by each sesor ode must ultimately be trasferred to a remotely base statio. The commuicatio is expesive i terms of eergy usage betwee sesor odes ad base statio. A fusio ode is selected i each WSN to commuicate with the remotely base statio o behalf of other sesor odes i the efficiet maer [6]. Oce a sesor ode is selected a fusio ode, its overall eergy cosumptio icreases sigificatly as it has to commuicate with all other sesor odes withi coverage of regio of iterest (ROI) as well with the remotely base statio. Therefore the process cotiues periodically ad i (5)
www.ijcsi.org http://dx.doi.org/10.20943/01201604.1115 13 each roud differet fusio ode selected to balace the eergy cosumptio throughout the WSN. Fusio ode selectio is a importat procedural step due to the exact locatio of sesor ode. If the selected fusio ode is located closer to most of the sesor odes withi ROI, the cost of commuicatio betwee fusio ode ad other odes will be miimal. O the other had if fusio ode is located far away from majority of the sesor odes the the trasmissio cost from each sesor ode to fusio ode will be higher ad the other sesor odes will cosume more eergy as stroger sigal will be required to commuicate [7]. The probability of each sesor ode becomig a fusio ode based o its probabilistic distace from remaiig sesor odes is computed ad the oe with the highest probability is selected as the fusio ode. Sice the exact locatio which is required for fusio ode selectio is ot kow, therefore each ode calculates the distace probabilistically. The distace depeds upo the iformatio a ode receives from all other ode i the WSN. This iformatio comprises of the sigal stregth ad eergy. Bayesia etwork is used to calculate the probabilistic distace betwee each pair of odes. The probabilistic distace together with residual eergy of each ode eables the Bayesia etwork to fid the most probable cadidate to become a fusio ode. The whole process is repeated to obtai ew fusio ode after each roud ad the residual eergy is adjusted accordigly after every roud. Sice a WSN possibly cosist of from tes to hudreds of sesor odes, size of WSN depeds upo type of applicatio where the etwork is employed. I this paper as we are focusig oly o the fusio ode selectio process, a small BAN is used that cosist of five odes scattered radomly i a area. Figure 1 shows the sceario used as the example. The figure shows a sceario where five sesor odes are located i a BAN ad are scattered radomly. From the Figure 1 it is evidet that sesor ode umber 5 is located far from the rest of the four other odes. ode 1, 2 ad 3 are closed to each other ad ode 4 ot far from ode 2 ad 3. I our research sice locatio of sesig odes are ot kow to the system, therefore we have employed Bayesia Network to fid the probabilistic distace amog the odes i the WSN. ode 1 ode 3 ode 2 Target Reportig Chaels ode 4 ode 5 Figure 1 The odes scattered radomly i a ROI of WSN 4. Data miig discovery fusio ode ROI of WSN I this sectio, we formulate Bayesia Network model to discover the fusio ode. At the start of each roud, each sesor ode ca potetially become a fusio ode, the selectio depeds upo the locatio of the ode with respect to rest of the sesor odes i the ROI of WSN ad the residual eergy level. Therefore at the start of each roud each ode will calculate its distace probabilistically from all other sesor ode i the WSN, the distace depeds upo sigal stregth ad power level. As every sesor ode i the WSN becomes a fusio ode therefore Bayesia Network must be built for every ode cosiderig it to be a cadidate for fusio ode. A example of such etwork for ode N1 is show i Figure 2 below; Figure 2 Bayesia Network for ode 1.
www.ijcsi.org http://dx.doi.org/10.20943/01201604.1115 14 Figure 3 below shows that how sesor ode 1 sees the other sesor odes i the WSN from its ow perspective. H1 is the probability of sesor ode N1 to become a fusio ode, this probability is calculated by ruig the simulatio usig Netica @. D1 is the probabilistic distace that shows how close ode N1 is from all other odes withi the WSN. D21 is the distace from sesor ode N2 to N1, D31 is the distace from sesor ode N3 to ode N1 ad so o. S21 is the sigal stregth received by sesor ode N1. P1 is the power level of sesor ode N1 ad so o. Probabilistic calculatio of D21 is based o the sigal stregth S21 ad power level P2. The same model is repeated for every ode i the WSN. Idividual odes will receive probabilistic distace from every other ode ad the calculates the overall aggregate distace from itself to all other sesor ode i the WSN. (2) All sesig odes are immobile; (3) All sesig odes are homogeous, ad are eergy costraied; (4) odes have o locatio iformatio; (5) After every roud each other odes will cosume 5% eergy; (6) Every fusio ode after each roud will cosume 8% of eergy. ode N5 i the preset sceario is located comparatively far from the rest of the sesor odes therefore it cosume more eergy, 6% of eergy will be cosumed by sesor ode N5 after each roud. Figure 4 below shows the result of the fifth simulatio for sesor ode N1, it ca be see from above that i the preset sceario with the available sigal stregth ad power level, ode 1 has 72.8% probability of becomig a fusio ode (H1) for the fifth roud. I the same way each ode will calculate its ow probability of becomig a fusio ode, ode showig highest probability will become fusio ode head for the curret roud. At the ed of each roud the ode showig the highest probability amog the five sesor odes will become a fusio ode for that particular roud. Figure 3 Simulatio result from Netica showig probability of sesor ode N4 which was selected as fusio ode H4 i the first simulatio ru Figure 3 above shows the result of simulatio for sesor ode N4, it ca be see from above that i the preset sceario with the available sigal stregth ad power level, H4 has 85.7 % probability of becomig a fusio ode for the preset roud. I the same way each ode will calculate its ow probability of becomig a fusio ode, ode showig highest probability will become fusio ode for the curret roud. I our simulatio followig assumptios have bee made: WSN coverages are already formed before the fusio ode selectio process ad the WSN s ROI could be of differet size. The assumptios list as the followig: (1) The remotely base statio is located far from the WSN coverage; Figure 4 Simulatio result from Netica showig probability of sesor ode N1 which was selected as fusio ode H1 i the fifth simulatio ru Table 1 below shows the simulatio results, at the start of roud 1, each sesor ode carries 98% eergy, from P1 to P5. The probability of each ode is listed from H1 to H5. At the ed of each roud, ew power level value is calculated for ext roud, reduced power level is recorded ad ew fusio ode is selected accordigly. As ca be see from the etry i the table that for roud 1,
www.ijcsi.org http://dx.doi.org/10.20943/01201604.1115 15 ode H4 has the highest probability amog the five odes ad hece selected as fusio ode for roud 1. Table 1 Ro u ds Simulatio results from Netica usig Bayesia Network Approach P1 P2 P3 P4 P5 H1 H2 H3 H4 H5 Cadidate Fusio o de 1 0.98 0.98 0.98 0.98 0.98 0.82 0.82 0.84 0.86 0.77 H4 2 0.93 0.93 0.93 0.90 0.92 0.79 0.80 0.82 0.82 0.75 H3 or H4 3 0.88 0.88 0.88 0.82 0.86 0.77 0.78 0.80 0.79 0.73 H3 [4] Z. Ha, R. Zheg ad H. V. Poor, Repeated auctios with Bayesia oparametric learig for spectrum access i cogitive radio etworks, Wireless Commuicatios, IEEE Trasactios o, vol. 10, o. 3, 2011, pp. 890-900. [5] X. Xig, T. Jig ad Y. Huo, Chael quality predictio based o Bayesia iferece i cogitive radio etworks, INFOCOM, Proceedigs IEEE, 2013, pp. 1465-1473. [6] Bill C.P. Lau, Ede W.M. Ma, Tommy W.S.Chow, Probabilistic fault detector for Wireless Network, Expert System With Applicatios, Volume 41, Issue 8,15, 2014, pp 3703-3711. [7] X. Xig, T. Jig ad W. Cheg, Spectrum predictio i cogitive radio etworks, Wireless Commuicatios, IEEE, vol. 20, o. 2, 2013, pp. 90-96. 4 0.83 0.83 0.80 0.77 0.80 0.74 0.75 0.72 0.73 0.71 H2 5 0.75 0.78 0.75 0.72 0.74 0.73 0.70 0.69 0.69 0.70 H1 5. Coclusios I this survey paper, we address the problems of fusio ode selectio i wireless sesor etwork. Due to the wireless sesor etworks is large, ad it correspods to differet time ad place, the record of the spectral iformatio of data is very huge, i such big data coditio. Ucertaity is a iheret property of WSN Sesed data. Uder the architecture of Bayesia etwork, which ca itegrate the quatitative ad qualitative kowledge ito a comprehesive probabilistic kowledge represetatio ad iferece eviromet, this paper presets a BN model for data miig from sesig data. Ackowledgmets This research work was sposored by the Miistry of Sciece ad Techology, R.O.C., uder project umber MOST104-2221-E-155-019. Refereces [1] Y. Gai, B. Krishamachari ad R. Jai, Learig multiuser chael allocatios i cogitive radio etworks: A combiatorial multi-armed badit formulatio, New Frotiers i Dyamic Spectrum, IEEE Symposium o IEEE, 2010, pp. 1-9. [2] Davood Izadi, H. Jemal, Sara Ghaavati Abawajy, ad Herawa Tutut. "A Data Fusio Method iwireless Networks." s, vol.15, o. 2 (2015), 2964-2979. [3] Yee Mig Che ad We-Yua Wu Cooperative Electroic Attack for Groups of Umaed Air Vehicles based o Multiaget Simulatio ad Evaluatio, Iteratioal Joural of Computer Sciece Issues, vol. 9 (2), 2012, pp. 11 16.