Novel Approach for Detection and Tracking of Explosive Carrying Mobile Humans with Odor-Sensor Based Multisensor Networks

Transcription

1 Novel Approach for Detecton and Trackng of Eplosve Carryng Moble Humans wth Odor-Sensor Based Multsensor Networks Ahmet KUZU Control Engneerng TUBITAK-MAM-BTE PK1 17 Gebze/Kocael TURKEY METIN GOKASAN Electrcal Electroncs Eng. Fac. Istanbul Techncal Unversty Maslak 9, Istanbul, TURKEY Seta BOGOSYAN Electrcal Computer Engneerng Unversty of Alaska Farbanks Duckerng Buldng 1 Farbanks AK USA ffsob@uaf.edu Abstract: - In ths study, a dstrbuted system whch could be used to detect and track an eplosve carryng moble human s desgned. As a novel approach to estng odor detecton and trackng methods, a technque whch combnes odor and sound detecton s proposed for ncreased accuracy. To ths am, a MEMS sensor network ncludng a pressure sensor, an accelerometer, a mcrophone, an odor sensor and electronc unt s proposed. For the electronc unt of ths system a sensor network and least square estmaton based sensor fuson algorthm s developed and smulated. The results are evaluated wth respect to varous paths of two humans, one carryng the eplosve. The results show that odor sensng alone s not suffcent for the accurate determnaton of the odor track, and that by addng other conventonal trackng methods, accuracy could be ncreased. Key-Words: - Odor, Trackng, Sensor networks, Sensor fuson, Least squares, Eplosve 1 - Introducton Although there s suffcent technology to detect metal weapons, detecton of plastc eplosves such as C wth methods that are harmless to humans, stll remans to be a challenge. Metal detectors or -ray facltes commonly used at arports and customhouses cannot fnd plastc eplosves. Y-ray detecton usng rradaton of neutron beams s recognzed as an effectve method for C detecton at ths pont, but few actually use ths method due to doubts about the safety of usng nuclear radaton on the publc. Hence, methods must be sought to carry out C nspecton safely, nondestructvely, and wthout physcal contact. One possble method s the use of THz radaton, whch s clamed not to harm the human body n [1]. Detecton of eplosves usng odor sensors s already n use for the detecton of land mnes and there has been ongong research on odor sensor based detecton of plastc eplosves [][],[],[5], whch mostly develop ways to detect the dfferent odorants n a specfc odor. There are numerous studes n the lterature for odor source localzaton; some studes try to locate the odor source by trangulaton [], whle some studes try stochastc estmaton technques lke Least Square Estmaton Method [] and Mamum Lkelhood Method [7], or Bayesan based methods lke PQS (Process Query System) []. A common characterstc n all these algorthms s ther low accuracy and senstvty to envronmental effects lke the wnd. The major contrbuton of ths study s the development of a method that could be used wth odorsensor based mcroelectromechancal (MEM) sensor networks for the detecton and trackng of a human carryng a plastc eplosve whle movng among several other moble, unarmed humans. In ths study, t s also demonstrated that combnng odor detecton wth sound detecton wll provde hgher accuracy than fndng the odor source drectly, wth the use of odor detecton alone. In the developed algorthm, commonly used passve target trackng technques such as acoustc, sesmc and pressure based technques are fused to accurately detect and track the target trajectory. These targets are then matched wth the odor detecton results, whch are obtaned va the conventonal trangulaton method. The resultng performance provdes a hgher accuracy n comparson to other odor source localzaton technques reported n the lterature Another contrbuton of the study s the detecton and trackng of moble odor sources (.e eplosves n ths case) wth the use of statonary sensors, unlke most studes n whch statonary targets are pursued usng moble sensors [9]. Ths approach wll also ncrease the chances of trackng and dsarmng the ntruder wthout hs/her awareness and at an approprate moment and wll sgnfcantly lmt the harm caused to nnocent bystanders and securty offcers. The use of odor sensors wll further contrbute to optmze energy consumpton n the network n that all sensor groups n the module wll be kept on standby, unless a command s receved from the odor sensor. Issue, Volume, February

2 Ths study wll combne these approaches for the am of accurate detecton and trackng of a human carryng an eplosve n a publc place. To ths am, a multsensor network (MSN) wll be desgned based on n-mem sensor modules; each MEM sensor module wll nclude low-cost, low-power passve sensors, namely odor sensors for the detecton of the eplosve, and a mcrophone, accelerometer, passve nfrared sensors (PIRs) and a pressure sensor for the trackng of the human carryng the eplosve. Each MEM module wll also nclude crcutry for the data fuson of the odor sensors, sgnal condtonng crcutry as well as RF communcaton and power supply crcutry n the same case. Sgnal condtoners wll convert sensor outputs n each module to approprate levels, whch are then processed va analog-to-dgtal converters (ADCs). Data from the odor sensors wll also undergo a fuson process. The fused data wll then be transmtted from each module, along wth the outputs of the multple sensors to the hub unt, va RF transmsson. Ths unt s the major processng unt located n each subdvson of the network and performs fuson and decson tasks, whle also conductng communcaton wth each sensor module, other hub unts n the network and the base computer for hgher level data fuson and decsonmakng. Ths process wll be further dscussed n the Proposed multsensor networkng and sensor/data fuson approaches secton. A functonal block dagram of the process s gven n Fgure 1. Mcrophone Pressure Accelerometer ODOR 1 ODOR ODOR Mcrophone Pressure m Accelerometer ODOR 1 ODOR ODOR Mcrophone Pressure m Accelerometer ODOR 1 ODOR ODOR m Data Fuson Data Fuson Data Fuson RF Comm SENSOR MODULE 1 RF Comm SENSOR MODULE RF Comm SENSOR MODULE HUB UNIT SENSOR FUSION Fgure 1. Functonal block dagram of each MEM sensor module and hub unt Four dfferent sensor types are consdered for each MEM sensor module; namely, odor sensor, mcrophone, n accelerometer [], and pressure sensor. Structural dagram of ths mechatronc module s shown n Fgure 1. Odor sensors detect the eplosves whle mcrophones serve as passve radars, whch detect the locaton of moble objects from ther vbratons n the ar. Accelerometers, whch detect sesmc waves, also serve the same purpose. Fnally, pressure sensors are actvated upon drect contact, n that sense; they produce hghest accuracy and hence, are also used to calbrate the other sensors. The Zgbee Network protocol [11], n RF Communcaton Block, s used to collect the sensor data, whch are then used n fuson and decson algorthms [1]. The Rf Communcaton block also supples sensor localzaton by usng tranglzaton method [1,1]. Envronment Mechancal Part Odor Sensors Pressure Snsor Mcrophone Accelerometer ElectrcalInformatc Part Analog Analog Scalng Flterng Analog Scalng Flterng Analog Scalng Flterng Analog Scalng Flterng Fgure : Structural dagram of ths mechatronc module. The odor sensor s comprsed of a sensor array [15] to measure the densty of a varety of odorants consttutng an odor, a classfer to fuse and classfy the collected data [1,17]; Ths sensor detects the odor of weapon, and estmate the source of odor. [1]. Although the resoluton of odor sensors s lower than the other sensors, ther use n combnaton wth the odor sensor for the appromate localzaton of the target proves to be very effectve as demonstrated by smulatons n ths study. The followng sensors s desgned ndvdually and then, combned n a sngle case to consttute the above mentoned MEM sensor module: Sound sensors (MEM mcrophones): A mcrophone s a sensor, whch converts acoustc pressure to voltage by usng a thn daphragm. When voce pressure strkes a daphragm, stress and depleton occurs. Ths stress can be easly measured by pezoresstors, or depleton can be detected by capactance change. Fgure depcts the archtecture of a pezoresstve mcrophone. Sgnals output by the mcrophone provde some nformaton on the poston of the ntruder based on the duraton of tme requred for the ntruder-transmtted sound waves to arrve at the mcrophone. Ths data component wll be further fused wth data produced by other sensor modules at dfferent locatons to yeld the poston of the ntruder. The need for other sensors n addton to the Analog Swtch ADC Dgtal RF Mult Sensor Fuson Communcaton RF Modulaton Power Amplfers Sensor Network Issue, Volume, February

3 sound sensor arses from the fact that sound waves are subject to losses and reflecton; relable detecton and trackng requres the ablty to flter out these effects. pezoresstor mass daphragm pezoresstor Fgure. MEM Pezoresstve Mcrophone. Pressure sensors: Pressure sensors, gven Fgure, operate on the same prncples as mcrophones. The man dfference s the strength and resonance frequency of the daphragm. The pressure sensors operate as logc sensors, gvng the eact poston of the ntruder (f stepped upon), hence servng as a calbrator for the other sensors. gas capactance daphragm Odor Sensors: These sensors operate based on odorants. Each odorant compound sgnfes a dfferent odor. Typcal ways of measurng odor s by desgnng flters for the desred odorants, and placng a sensor at the output of each flter (see Fgure for a demonstraton of ths process). For nstance, f a capactve sensor s used, the capactance value vares based on the acton of these odorants, whch affects the delectrc part of the capactance, as demonstrated n Fgure 7. Each odor conssts of one or more odorant types. For nstance Odor 1 can consst of % Odorant A and 77-79% Odorant B. Another odor may consst of 5-55% of Odorant A and -5% Odorant B. Hence, recognzng an odor requres more than one sensor and the fuson of ther outputs. Senstve Layer Transducer Sgnal Data Processng Fgure. Odor Measurement Process. OUT bulk Fgure. MEM Capactve Pressure Sensor ntegrated capactor bulk Accelerometers: When acceleraton occurs, proof mass apples force on the beam, gvng rse to stress and depleton. Ths stress can be easly measured by usng pezoresstors, or depleton can be detected by ts capactance change. The archtecture of a MEM pezoresstve accelerometer appears n Fgure 5, whch demonstrates how ths system wll be used to sense moton n a certan area va vbratons on the ground. The PI and her PhD student have already desgned a MEM multple accelerometer module, also ncludng the data fuson and RF crcutry n the same case. The product s currently under testng. pezoresstor pezoresstor Slcon sland on membrane Fgure 7. MEM Odor Sensor. There are bascally fve dfferent odor sensng methods: Conductvty Sensors, Pezoelectrc Sensors, Metal-ode-slcon feld-effect-transstor (MOSFET), Optcal Fber Sensors, and Spectrometry-Based Sensors. Methods for odorant sensng under consderaton n ths study are the followng: One method s measurng the change n the delectrc coeffcent of a certan polymer whch reacts wth the certan odorant. The delectrc coeffcent can be easly measured by makng a capactance usng ths polymer. The proposed structure s gven n Fgure. beam Fgure 5. MEM Accelerometer. proof mass 5 Issue, Volume, February

4 Metal Polyner Substrate Metal R odorant C odorant Fgure. Structure of resstve and capactve sensors and cross-sectonal vew The second approach s measurng the resstance of a certan polymer whch reacts wth the certan odorant. By makng a resstance usng ths polymer, the conductvty coeffcent whch changes as a result of the reacton can be measured. The structure s smlar to the capactance method, but a dfferent polymer s nvolved. The thrd approach s measurng the chemcal propertes (lke ph) of a specal polymer to dentfy the odorant. Once agan the approprate polymer reactng to the odorant s chosen, but n ths approach, the change n the chemcal propertes s measured by a chemcal FET. The source dran current of the FET s affected by the chemcal propertes of the polymer. The structure of the chemcal FET s gven n Fgure 9. Polyner consumpton, the fused data n each SM s transferred to the hub unt va the neghborng SMs. Hence, n the 1st year, the dutes of each SM nvolve sensng moton and odor, detectng odor and passng on ts data and data from other SMs to the hub unt. s (HU): These unts possess a hgher processng power and memory n comparson to the sensor modules and conduct the fuson of data comng from sensor modules n ts subdvson. HUs are also equpped wth power and wreless communcaton crcutry. Each HU s n charge of a certan subdvson (ndcated by the crcles n Fgure ), hence, collects data comng from SMs related to the locaton of the target and odor of the eplosve to further perform data fuson and decson-makng for the proper determnaton of the moble eplosve s moton trajectory. Ths nformaton s then transmtted to the neghborng HUs and the base. For optmzed energy consumpton, the transmsson of the data to the base s also done va other HUs, where all data s fused for decson-makng. Metal Source Substrate Metal Dran Fgure 9. Chemcal FET V(gs) odorant The desgn of the MEM odor sensors n ths study requres for a hybrd structure. Snce the odor detecton wll requre several odor sensors to be ncorporated n the same module, wth each sensor reactng to a dfferent attrbute of the eplosve s odor, a data fuson algorthm takng place nsde the module wll also be developed to fuse the data from the odor sensor array. For energy optmzaton purposes, the sensors and fuson/communcaton process n each module and n the sensor network wll be actvated only when the odor sensor n a sensor module ndcates a potental eplosve n that subdvson. Addtonally, to mnmze the energy Fgure. Sample confguraton of the MEM-MSN - Physcal Model for Intruder Detecton and Trackng usng Sound Sensors and Accelerometers Used In Smulaton Fgure 11 gves a confguraton eample for the random locatons of the sensor modules, hub unt and the eplosve carryng human (ntruder). As wll be demonstrated wth the derved equatons, the locaton of the ntruder s determned wth a mnmum of sound sensors, denoted by S 1, S, and S, the data from whch s fused at the hub unt va Least Square Estmaton. The hub unt n that sense s assumed to be just another sound sensor, whch hears the ntruder as soon as any one of the sensors sends t a sgnal. In the smulatons, S1 s assumed to have heard the ntruder frst; hence, both S1 and hub unt are assumed to hear the ntruder n dt seconds. As can be seen n Fgure 1, the unknown Issue, Volume, February

5 dstance between the hub unt and the ntruder s denoted by d. The dervaton of the equatons are based on the notaton used n Fgure 1. The coordnate aes of the hub unt s taken as reference;,y are the coordnates of the ntruder, whle ( 1,y 1 ), (, y ) and (, y ) are the coordnates of S 1, S and S, respectvely. Sound s a travelng wave of pressure and can be modeled as, f ( d, v, t) = K( t) f ( t) (1) where f(t) s the sound sgnal and K(t) s an ampltude coeffcent whch decreases n tme; v s the velocty of sound, whch s appromately m/s n the ar at C. At an arbtrary nstant, t, an ntruder appears wthn the range of S 1 at a dstance, d 1 from S 1 as demonstrated n Fgure. At dfferent tme nstants, S and S S 5 also hear the ntruder, wth d denotng the dstance between S and ntruder, d denotng the dstance between S and ntruder, and so on... dt denotes the arrval tme of the sound at S 1 and hence, at the hub unt. dt 1 denotes the arrval tme dfferences between S 1 and S ; dt denotes the arrval tme dfferences between S 1 and S. Note that dt s unknown, whle dt 1, dt are measured. y Y S f(t) Intruder dt d 1 d dt 1 dt Sensor #1 d Sensor # Sensor # Fgure 1. Intruder ntated sound waves wth respect to sensors and hub unt Base on these defntons, the followng relatonshps can be derved: d 1 = v( dt ), d = v( dt + dt1 ), d ( ) = v dt + dt () Usng (), the followng relatonshps can be gven between each sensor and ntruder coordnates: ( = dt 1 ) + ( y y1) ( v( )) () ) + ( y y ) = ( v( dt 1 )) () ) + ( y y ) = ( v( dt )) (5) ( + dt ( + dt t t t t y y 1 y d 1 S 1 HUB UNIT d d In trud er d 1 Fgure 11. Confguraton and defntons for ntruder, sensors and hub unt S Combnng () wth () we get (), () wth (5) we get (7), and () wth (5) we get (). As can be seen from the equatons,,y, and dt are the unknown varables, whle all terms on the rght are known. Hence, t can be seen that determnng the ntruder s locaton n -dmensonal spece requres sensors, whle - dmensonal space would requre a mnmum of sensors. At ths pont, to reflect the more realstc nature of sound velocty, v, whch s actually dependent on the materal and temperature, a whte nose s added to the above equatons and the soluton s sought n terms of a Least Square Estmaton problem. ( 1 ) + ( y y1) y + v ( dt1 ) dt = v dt1 y y1 ( ) + ( y y ) y + v ( dt dt1 ) dt = v dt1 + v dt ( 1 ) + ( y1 y) y + v ( dt ) dt = v dt 1 y1 + + y () y + + y (7) () ( 1 ) + ( y y1) y + v ( dt1 ) dt + ε1 = v dt1 + + y 1 y1 (9) ( ) + ( y y ) y + v ( dt dt1 ) dt + ε = v dt1 + v dt + + y y () ( 1 ) + ( y1 y ) y + v ( dt ) dt + ε = v dt 1 y1 + + y (11) Issue, Volume, February 7

6 - Physcal Model for Intruder Detecton and Trackng usng Pressure Sensors In Smulaton Each module has a pressure sensor operatng as logc sensors. In the eperments, they wll be actvated upon steppng and hence, serve as a calbrator for the other sensors and detecton/trackng algorthm; however, n the smulatons, these sensors are assumed to be actvated when the ntruder comes wthn a very close range of that certan pont representng the sensor module. For eample, n the presented smulaton results, only S 5 (sensor 5) was stepped on. - Prelmnary model for odor dffuson and measurement n the smulatons Further assumptons n the smulatons are that the eplosves are emttng a gas constantly, whch gves rse to a certan gas densty at every nstant, t, and ths gas dsperses nto the envronment by dffuson. Hence, a gas startng to dsperse nto the envronment at t 1, wll lose ts effect n tme but wll est forever. From the followng nstant, t and on, new gas amounts wll be effectve cumulatvely at any gven pont (or sensor module) n the envronment.. The new gas amount wll also dsperse nto the envronment va dffuson, but wth a reduced effect as tme advances. Each odor sensor senses and measures the cumulatve gas amount at that pont. Snce Fck s nd law s a partal dfferental equaton wth only a few analytcal solutons avalable for specal cases, numercal methods have to be used to apply t to techncal problems. An analytcal soluton of Fck s nd law s gven by Equaton (1): [19] d c( t, d) = c Γ. erfc (1) Dt where c: dspersed gas amount, c Γ : gas amount at source, d: dstance from source, D: constant. In the smulatons, odor s consdered as a gas dspersng n amounts of c at every t. As the eplosve-carryng human moves, at every pont he/she moves at an new nstant t, a new gas source, c Γ starts ts dsperson based on (1). A gven odor sensor n the envronment located at a dstance, d from the ntruder wll accumulate n gas amounts, from n c Γ sources, gvng rse to a gas amount, S as epressed below: S = n k = 1 c ( t, d) (1) k n the smulatons. Hence, for the determnaton of the coordnates for the eplosve-carryng ntruder, a weghted-average s performed combnng the measured S 1, S, S,...amounts collected from all odor sensors (5 n ths case) at coordnates 1, y 1,, y,, y... to determne the eplosve-carryng ntruder s coordnates, and y. SY y =, S S = (1) S As the smulaton also consders more than one humans (wthout eplosves) n the envronment, the dstncton of a dfferent human s made by evaluatng the traveled dstance n the epred tme nterval n terms of ts feasblty for a human-beng. The same approach s appled for the determnaton of coordnates usng the sound and pressure sensors. - Smulaton Envronment In the smulaton, a hub unt and randomly deployed ffty sensor modules are consdered n a by square meter area. The area s assumed to have no ar draft, no boundares, hence no reflecton effects. The envronment temperature and the velocty of sound throughout the smulatons are assumed to be constant. It s also assumed that the humans (one carryng eplosves) are emttng sound waves perodcally. At ths pont of the studes, t s also assumed that the locaton of each sensor module and hub unt s known. Each module n the 5-sensor network conssts of an odor sensor, mcrophone, accelerometer and pressure sensor. The total smulaton tme s 5 seconds, wth the samplng tmes for sound/pressure/ accelerometers taken as.1 sec and that of odor sensors (wth a slower response) taken as 1sec. In these results, data fuson s performed usng a Least Square Estmaton algorthm, whch fuses all odor sgnals and moton sgnals separately combnng data comng from sensors of the same type. However, a decson-makng process whch evaluates odor and moton sgnals together s also performed n determnng the moton of the moble eplosve and separatng t from the trajectory of the unarmed human. In the smulaton results, the moton of sound and odor dffuson s taken nto consderaton wth the followng models: Although a fuson and ANN based decson algorthm wll be conducted for the multple odor sensors n one module, only one odor sensor per module s consdered Issue, Volume, February

7 5 - Smulaton Results The smulaton results for the detecton and trackng of the eplosve-carryng human (n an envronment where an nnocent human s also walkng around) are gven n Fgure 1. In the fgures, the * s ndcate the locaton of the sensor modules; the straght lnes ndcate the actual trajectores of the humans n the envronment; the dark grey trangles represent the estmated trajectory of the nnocent human and lght grey trangles represents that of the eplosve-carryng human; fnally the black crcles ndcate the trajectory of the moble eplosve estmated based on data collected from the odor sensors only. In Fgure 1a we see two people walkng parallel n north east drecton. People at the south has bomb. It s easly seen that algorthm fnd the correct person easly. In Fgure 1b we see two people walkng n parabolc trajectory from west to east. And the person at the north has bomb. In the fgure we can predct the person who has the bomb. And we also see that we should trace the peoples before we sense the odor and use these data. Because the odor dffuse too slowly. In Fgure 1c we smulate two person one of them s walkng to north east drectly, and other one s walkng on a spral trajectory of ncreasng radus. And n Fgure 1d both two people walkng on the same knd of spral trajectory. As t s easly seen that, the bomb can be easly detected agan n these case. In Fgure 1e one person s walkng to south west drectly, and other one s walkng to south east drectly. And ther trajectores are ntersected. As seen n ths Fgure ths case can also be determned. And n Fgure 1f the people s walkng on parabolc trajectory, whch are ntersected. As t s seen n the fgure determnng the person who has bomb get much more complcated, but stll t can be done. Man problem s the speed of odor wth respect to the ntruder. So other sensors are gettng a necessty. The proposed odor sensor-based network actvaton strategy, whch wll obvously yeld reduced energy consumpton, s also analyzed n terms of mean square error(mse) of trackng accuracy. Wth a tral smulaton run usng ths strategy, only sensors became actve to track the moble eplosve. The strategy resulted n an MSE of.5 as opposed to the conventonal strategy keepng all 5 sensors n the network, yeldng an MSE of.. The very mnor accuracy drop wth ths smplstc smulaton run motvates the use of the proposed strategy n the actual network for energy optmzaton. - Conclusons Odor dffuses very slowly so we can easly say that, although accelerometers and sound sensors are more effectve n trackng due to ther fast response, odor sensors are effectve n detectng the odor, but not n trackng due to ther slow response. Moreover, n a nosy envronment, accelerometers could effectvely assst the sound sensors. Nonunformtes of the ground are very effectve on the performance of accelerometers; hence, sound sensors could be more effcent n such cases. Velocty of sound depends on ambent temperature, hence could beneft from calbraton. Pressure sensors could serve ths purpose. Pressure sensors have hgh accuracy, but low possblty of beng stepped on. Therefore we can say that sensor fuson s needed supplyng system more accurate and robust. Wth the developed algorthm, t s possble to detect and track the target usng data gathered from a mnmum sensors of the same type. However, usng data also from dfferent sensor types helps dstngushng the 1a) 1b) c) 1d) e) 1f) Fgure 1. Smulaton of the system by varous trajectores. Wth nformaton obtaned by fusng the data from each dfferent sensor type, t s possble to eclude data Issue, Volume, February

8 that s not n harmony wth data collected on the target. Hence, fusng the data comng from each sensor n the MEMS sensor modules wth a scalng factor helps ncrease target detecton and trackng. The evaluaton of the multple sensor data n relaton to the odor sensor data s essental n the accurate trackng of the moble eplosve n ths study. Odor sensors are not only capable of the ntal detecton of the eplosve, but also capable of servng as a somewhat slow gude n relatng the detected odor to the possble trajectores wth the help of the other sensors, as observed n the smulaton results. Logc data comng from pressure sensors wll act as a calbrator for poston determnaton and contrbute to accuracy. Odor sensor based network actvaton may possbly cause a small decrease n trackng accuracy, hence should be consdered for reduced energy consumpton n the network Thus, the above combnaton of sensors wll be consdered n the prelmnary desgn process of the MEM sensor module. Whle desgnng the MEM odor sensors, the decson on the choce of the supportng moton sensors wll also be fne-tuned wth more detaled and realstc smulaton tasks gven as student and/or master thess projects. References: [1] Yamamoto, K., Yamaguch, M., Myamaru, F., et. all., Inspecton of C mal bomb by terahertz spectroscopy, Conference on Lasers and Electro-Optcs (CLEO), Vol.1,, pp.. [] Shelby F. Shuler, Detectng Eplosves usng Electronc Noses, The Unversty of Alabama, Student Semnars,. [] T. Nakamoto, A. Fukunda, and T. Morzum. Perfume and Flavor Identfcaton by Odor-Sensng System Usng Quartz-Resonator Sensor Array and Neural Network Pattern Recognton, Actuator B, Vol.1/19, 199, pp. -9. [] Ross J. Harper, Jos e R. Almrall, Kenneth G. Furton, Identfcaton of domnant odor chemcals emanatng from eplosves for use n developng optmal tranng ad combnatons and mmcs for canne detecton, Elsever Talanta, Vol.7, 5, pp [5] K.Langer, A Gude to Sensor Desgn for Land Mne Detecton, IEEE, EUREL, 199, pp.. [] M.P. Mchaeldes, C.G. Panayotou., Plume Source Poston Estmaton Usng Sensor Networks, Proceedngs of the 1th Medterranean Conference on Control and Automaton Lmassol, Cyprus, 5. [7]. Kuang, H. Shao, Mamum Lkelhood Localzaton Algorthm Usng Wreless Sensor Networks, Proceedngs of the Frst Internatonal Conference on Innovatve Computng, Informaton and Control (ICICIC'),. [] G. Nofsnger, G. Cybenko, Dstrbuted Chemcal Plume Process Detecton, MILCOM, 5. [9] W. Jatmko, T. Fukuda, T. Matsuno, F. Ara, B. Kusumoputro, Robotc applcatons for odor-sensng technology: Progress and challenge, WSEAS Transactons on Systems, Vol., No.7, 5, pp [] K. E. Speller, D. Yu, A low-nose MEMS accelerometer for Unattended Ground Sensor applcatons, Proc. SPIE, Vol.517,, pp.. [11] Evans-Pughe, C, Bzzzz zzz [ZgBee wreless standard], IEE Revew, Vol.9,, No., pp.-1. [1] F. Zhao, T. Gubas. Wreless Sensor Networks: An Informaton Processng Approach, Elsever, [1] K.C. Santos, E.F. Slva, J.N. Carvalho and H.C.C. Fernandes, Rado Localzaton n Moble Comuncatons Systems, WSEAS ISCA 5, pp.15 [1]. L, A.A. Ahmed, H. Sh, Y. Shang, A Systematc Evaluaton of RangeQ-based Localzaton Algorthms n Wreless Sensor Networks, th WSEAS Internatonal Conference on Crcuts, Systems, Electroncs, ControlSgnal Proc., pp [15] H. Sundgren, F. Wnqust and I. Lundstrom. Artfcal Neural Network and Statstcal Pattern Recognton Improve MOSFET Gas Sensor Array Calbraton, Tech. Dgest Transducers, 1991, pp.57. [1] P.M. Schwezer-Berberch, J. Goppert, et. al., Applcaton of Neural Network Systems to Dynamc Response of Polymer-Based Sensor Array, Actuator B, Vol./7, 1995, pp.-. [17] B. Kusumoputro, Structural and smoothng parameter optmzaton of probablstc neural network through evolutonary computaton and ts usage n odor recognton system, WSEAS Transactons on Computers, Vol., No., 5, pp.9-1 [1] P. Keller, Overvew of Electronc Nose Algorthms, Internatonal Jont Conference of Neural Networks (IJCNN'99), Washngton, USA, [19] Crank, J. 195, The mathematcs of dffuson. Oford Unversty Press. 7 Issue, Volume, February