Advanced Bio-Inspired Plausibility Checking in a Wireless Sensor Network Using Neuro-Immune Systems

Transcription

1 Fourth Internatonal Conference on Sensor Technologes and Applcatons Advanced Bo-Inspred Plausblty Checkng n a reless Sensor Network Usng Neuro-Immune Systems Autonomous Fault Dagnoss n an Intellgent Transportaton System Amr Jabbar Insttute for Mcro sensors, Actuators and Systems, Department of Electrcal Engneerng, Unversty of Bremen Bremen, Germany aabbar@msasun-bremende alter Lang Insttute for Mcro sensors, Actuators and Systems, Department of Electrcal Engneerng, Unversty of Bremen Bremen, Germany wlang@msasun-bremende Abstract Recent developments n wreless sensng technology lead to mplement advanced algorthms for dstrbuted data processng n varous applcatons; ntellgent transportaton system s one of the man applcatons of the advanced networked sensng technology to montor the envronmental condtons for controllng the qualty of the products To ensure the desred performance of a wreless sensor network, the relablty of the network records needs to be evaluated usng an effcent data processng algorthm In ths paper, a new applcaton of a bo-nspred technque s ntroduced for autonomous plausblty checkng n a wreless sensor network; at frst, an optmzed Neuro-mmune system s ntroduced and developed to predct the sensor records; then, performance of the proposed Neuro-mmune system s compared wth a neural network approxmaton mechansm A secondary algorthm evaluates the sensor records to check the plausblty of the records n the wreless sensor network The proposed data processng algorthm could serve n varous applcatons of wreless sensor networks Keywords-wreless sensor network; dstrbuted data processng; artfcal mmune system; artfcal neural network I INTRODUCTION Due to recent developments n networked sensng technology, wreless sensor networks are becomng more wdely used to control and montor the envronmental condtons n varous applcatons lke transportaton systems A sensor network ncludes many smart devces to perform a requred task dependng on type and capablty of the sensor nodes [] Dfferent archtectures could be establshed to desgn a sensor network ether n a centralzed or a decentralzed manner; the requred decsons could be locally made n dfferent layers of a wreless sensor network In each nterconnected sensor network, t s necessary to evaluate the relablty of the sensor nodes by processng the recorded data of each node as well as each cluster In general, plausblty of a wreless sensor network denotes the ablty of the system to correctly record and make the approprate decsons adherng to desred envronmental condtons to perform a requred task The plausblty could be affected by undesred events (lke faults) whch lead to any unwanted devaton of sensor network records There are dfferent technques to evaluate the performance of the wreless nodes ncludng modern and classcal technques []; artfcal neural network (ANN) s a knowledge based technque ncludng nonlnear mappng features and generalzaton whch makes t favorte for model-free data processng [] An optmzed neural network was mplemented on a wreless sensor network to approxmate and classfy the records for plausblty checkng [4] The advantages and dsadvantages were hghlghted n comparson wth the classcal approaches (lke Least Squares) [5][6] Artfcal mmune system features, whch are establshed on human bologcal mmune system to detect and elmnate the threats, could be ntegrated wth dfferent technques lke neural networks to ncrease the flexblty and accuracy of data processng [7] In ths paper, a wreless sensor network s used to record the envronmental parameters ncludng temperature by mote sensor nodes A novel bo-nspred technque for autonomous plausblty checkng s ntroduced and mplemented The proposed optmzed bo-nspred technque could serve n varous applcatons of wreless sensor networks Related works are studed n Secton II; then, the concept of autonomous plausblty checkng n a wreless sensor network s descrbed n Secton III The dea of Neurommune system s dscussed to desgn the data processng mechansm n Secton IV; fnally, the expermental results usng the proposed data processng technque s presented II RELATED ORK Dstrbuted data processng could be used for any local decson makng lke power savng or routng n a wreless sensor network to ncrease the overall performance [8][9] Thus, accuracy and energy effcency of the chosen data processng technque are taken nto consderaton [] Data processng archtecture conssts of two man archtectures ncludng data approxmaton and classfcaton Data approxmaton s establshed on ether lnear or nonlnear mappng of varous data; then, the approxmated data could be used ether for data fuson purposes [4] To approxmate the records n a wreless sensor network, neural network / $6 IEEE DOI 9/SENSORCOMM4 8

2 could be an approprate choce due to ts nonlnear mappng features, preferable to lnear approaches lke Least squares Also, dependng on the applcaton, dfferent neural networks are employed to use n data classfcaton [] Artfcal mmune system s another bo-nspred technque whch could be combned wth neural network for optmzaton Dasgupta compared artfcal neural network wth mmune systems [] De Castro and Von Zuben performed comparatve studes about artfcal mmune system and neural network [] They developed a growng Boolean compettve network usng mmune system features ncludng compettve learnng characterstcs Intalzng weghts of the multlayer feedforward neural networks was another attempt to combne a neural network wth mmune system [] Ths approach leads the neural network to converge to a local optma whch mproves the performance of the network n some applcatons Consderng the recent efforts n desgnng the neurommune systems to desgn the optmzed neural networks, a novel bo-nspred data processng technque s ntroduced n ths paper to check the plausblty of the records n a wreless sensor network; the expermental results are used to compare the accuracy of an optmzed backpropagaton technque wth the proposed neuro-mmune system Fgure shows a cluster of a wreless sensor network ncludng mote sensor nodes; data are sent and receved va Imote s CC4 rado, wth a processng frequency of 4 MHz (n ths research) at RF power of - dbm for all tests [4] A data processng platform observes the sensor records at each cluster to make an approprate decson when one of the sensor nodes devates from desred behavor Each cluster s montored usng a local data processng platform (DPP) and the relablty of each group of data processng platforms s evaluated usng a global DPP III AUTONOMOUS PLAUSIBILITY CHECKING IN IRELESS SENSOR NETORKS Many sensor nodes are used n a wreless sensor network to montor the envronmental condtons n a transportaton system It s necessary to evaluate the relablty of the records to keep the qualty of the products hgh durng transportaton Any abnormalty n a wreless sensor network s detected, solated and nvestgated n an autonomous wreless sensor network Some decsons are made locally to evaluate the records of each cluster Fgure Local and Global data processng archtectures Fgure Local Data processng at a cluster n a wreless sensor network Fgure llustrates the generalzed data processng archtecture nsde a truck; accordng to ths fgure, each three sensor nodes are observed by a local data processng platform n each zone Then, the records of the local data processng platforms (A, B, and C n zone and D, E, and F n zone ) are processed together to detect any abnormalty n the sensor network A reefer unt establshes the desred envronmental condtons by coolng down or warmng up the truck Fgure shows a two stage data processng platform ncludng a knowledge based data approxmaton as well as a classfcaton mechansm To perform data processng at each cluster ether locally or globally, the sensor records are approxmated by the Neuro-mmune system Approxmated data are compared to the actual records of the under approxmaton nodes by generatng socalled approxmaton resduals 9

3 wreless sensor network The Neuro-mmune system deals wth the tranng weghts and nput set to tran the network Input Set Immunologcal eght adustment Fgure Bo-nspred data processng n a wreless sensor network At second step, the relablty of the records s evaluated usng a classfcaton algorthm to detect any abnormalty n the wreless sensor network Sldng correlaton factor of the sensor nodes n each cluster s a complementary condton to classfy the records The Neuro-mmune system data processng archtecture s ntroduced n next secton for data approxmaton and classfcaton n a wreless sensor network IV NEURO-IMMUNE SYSTEM DESIGN The artfcal mmune system s derved from human bologcal mmune system to defend the body aganst the threats Rememberng the past encounters and recognton features are the man specfcatons of the artfcal mmune systems [7] The Bologcal mmune system protects the human body from nfecton usng prmary response to nvadng pathogens as well as a secondary response to remember past encounters Antbodes (Ab) and Antgens (Ag) are two types of molecules; Ab molecules (cell receptors) bnd to Ag (pathogenc mcroorgansms) for ther posteror elmnaton The Ab and Ag partcpate n mmune recognton between the bndng regon of the receptor and eptope [7] For representng the artfcal mmune systems, the antbody and antgen vectors are shown as: Ab = Ab, Ab,, Ab L Ag = Ag, Ag,, Ag L Also, the affnty functon whch denotes the degree of match between the mentoned vectors could be descrbed by ether Eucldan or Manhattan dstances; D = D = L = L ( Ab Ag ) (Eucldan dstance) () = Ab Ag (Manhattan dstance) () In ths paper, mmune system s used to desgn an optmzed neural network to mprove the data processng n a sk,, s sk,, s sk,, s k + X () () Φ () () Φ () Stored Input Sets Input Layer Hdden Layeres Fgure 4 Proposed Neuro-mmune archtecture Output Layer Fgure 4 shows the proposed Neuro-mmune archtecture; to desgn the network, dfferent parameters have been explored (such as number of layers and neurons) n order to obtan the Maxmum data approxmaton accuracy when the temperature of a sensor s approxmated usng three neghborng sensor nodes at each cluster [4][6] Two hdden layers are selected each ncludng four neurons wth () () sgmodal actvaton functons ( Φ and Φ ); the network () () layers are connected usng the weght vectors (,, () and ) Each weght vector at each layer () ncludes some weght elements; ( ) = =,, ( ) ( ) ( ) [ w, w, w, w ] ) ( 4 A Sldng Backpropagaton Sldng backpropagaton was developed usng the last sequental records of all sensor nodes at each cluster; to approxmate the records of each under approxmaton sensor node, the neural network s traned usng the records of the neghborng sensor nodes [4] There are four sensor nodes at each cluster, one as a data processng platform, to montor the cluster; to predct new values of each under approxmaton sensor node, last four sequental records of the neghborng sensor nodes are assumed as the approxmaton parameters (as tranng nput); they are mapped nto the last four sequental records of the under approxmaton sensor node (as tranng target) to tran the neural network For example, to approxmate the records of s (at nstance k + 4 ), last four sequental records of the 4 neghborng sensors are collected as the approxmaton parameters DPP ( s, s and s ) at nstances k, k +, k +, and k + ; Y

4 s X = s sk k k DPP s s DPP s (Tranng nput) The target vector denotes the last four sequental records of the under approxmaton node whch s s n ths example [ s s s s ] Y (Tranng target) = k In sldng backpropagaton algorthm, the ntal weghts are chosen randomly; then, the weghts are updated usng the gradent descent algorthm to mnmze the error functon (Er) between the desred and actual outputs of the network durng the tranng phase [4] Er = [D - Y ] () In (), D and Y refer to the desred and actual outputs of s, respectvely The weght varatons are proportonal to the negatve gradents of the error functon and current weghts where η s the learnng rate n (4); Δ () () = η Er (4) Also, the weght vectors are consequently updated accordng to the negatve gradents of the error functon at each nstance ( k ) by (5) () (k + ) = =,, () (k ) + Δ After tranng, by feedng new values of the approxmaton parameters, the new record of the under approxmaton node s predcted Ths procedure s appled to predct the records of each node n smlar way B Neuro-Immune System Bascally, adustng the weghts has maor mpact on the qualty and speed of the data approxmaton usng neural networks Adherng to nsuffcent local mnma whch leads to nadequate data approxmaton s a bg problem n tranng the neural networks [5] For ths purpose, mmune system s mplemented to ntalze the weghts at each layer of the network The developed approach s based on a smulated annealng algorthm (de Castro et al) [] Ths approach generates a set of weght vectors to reduce the lkelhood of the network to converge to a local optma The smulated annealng (SA) technque s derved from atomc dsplacement n lquds; the atomc poston n a gven lqud () (5) sample s calculated by a probablty factor where E and T are the confguraton energy and temperature P( ΔE E) = exp( ) T Δ (6) Each atomc dsplacement leads to an energy change ( Δ E ); when the energy change s postve, probablty of acceptng the atomc dsplacement s calculated by (6) [] To defne the energy functon n a data approxmaton archtecture, Eucldean dstance ( Ds ) s used to check the affnty of the weght vectors n (7); 4 Ds ( w, w ) = ( w w ) (7) k = here w, w are the connected weght vectors to dfferent neurons n a layer The energy s defned as the sum of the Eucldean dstances among all weght vectors (antbodes) Ths energy s mnmzed to adust the network weghts by (8); thus, the weghts are selected by the mmune system to generate the dstrbuted vectors nstead of beng ntalzed randomly 4 4 E = ( w, w ) (8) = = + Another contrbuton of the mmune system s to observe and deal wth the nput set to tran the neural network As mentoned before, the accuracy of the sldng backpropagaton s hghly dependant on the last four sequental records of each sensor node; sometmes due to the naccurate readngs of the sensors, the dfference between the approxmaton and the actual value, whch s called approxmaton resdual, ncreases As n the sldng backpropagaton technque, only the four sequental records at k, k +, k +, and k + are taken nto account for data approxmaton at k + 4, the naccurate readngs wll lead to naccurate data approxmatons for the next sequental approxmatons The other case s when the sensor records are not plausble due to fault occurrences; therefore, t s necessary to remove the naccurate or faulty sensor readngs from nput set for the next data approxmatons For ths purpose, the mmune system s utlzed; at each nstance (t), last eght sequental records (as antbodes) and the resultant approxmatons (as antgens) (at t-8,,t-, and t-) are stored The Eucldean dstance between the last eght approxmatons and related actual values are calculated and the trend of data approxmaton s observed hen the Eucldean dstance ncreases consderably, t shows that sensor records are not accurate enough to be nvolved n the new tranng set Therefore, the naccurate or faulty records are not longer taken nto account to generate the new nput k k

5 set for next approxmatons The proposed neuro-mmune algorthm s descrbed brefly n Fgure 5 Fgure 5 Neuro-mmune system for plausblty checkng At frst, the last four sequental records are used to generate the nput set The Neuro-mmune system adusts the weghts and after data approxmaton, the resduals are generated If the resdual value s hgh, the naccurate records won t be nvolved n the new nput set for the next data approxmaton; otherwse, by calculatng the Eucldean functon, the plausblty of records s nvestgated; ths ssue as well as the hgh and low terms wll be further explaned n the followng secton V EXPERIMENTAL RESULTS The proposed neuro-mmune algorthm was mplemented on a wreless sensor network ncludng Imote sensor nodes [4] to record and process data for plausblty checkng The reefer unt could cool down or warm up the truck accordng to the set ponts; some arbtrary set ponts are used to test the performance of the proposed mechansm to establsh the transportaton condton The chosen reefer unt set ponts are 4, 8,, 6, and C respectvely, each takes 6 mnutes; therefore, the test duraton s mnutes The average ambent temperature s about 8 C All temperature records n dfferent postons are affected by the reefer temperature Accordng to the Fgure, two man zones are consdered for local and global data processng To examne the proposed Neuro-mmune algorthm, the records of the sensors A, B, and C are processed at Zone as well as nodes D, E, and F whch are local data processng platforms at zone At frst, the neural network (ANN) was mplemented to approxmate the records at zone (for nodes A, B, and C) by the sldng backpropagaton whch uses the last four sequental records to tran the network [4] 5 A-Approxmaton Resdual -5 5 B-Approxmaton Resdual -5 5 C-Approxmaton Resdual -5 Fgure 6 Approxmaton resduals usng the sldng backpropagaton The approxmaton resduals whch are the nstantaneous dfferences between data approxmatons and the actual records are seen n Fgure 6 The results show that ANN approxmaton resdual lays wthn ±5 C when the accuracy of the data approxmaton at each nstance s hghly dependant on the last four sequental records of each sensor node The calculated approxmaton resdual usng the Neuro-mmune system s llustrated n Fgure 7 The results show that the Neuro-mmune system produces less approxmaton resdual n comparson to the sldng backpropagaton network (± C) 5 5 A-Approxmaton Resdual B-Approxmaton Resdual C-Approxmaton Resdual -5-5 Fgure 7 Approxmaton resduals usng the Neuro-mmune system For a more precse comparson, the average of the calculated approxmaton resdual s presented n Fgure 8 at zone ; t compares the average approxmaton resdual of nodes A, B, and C usng the sldng backpropagaton (ANN) and Neuro-mmune system 5 Average-ANN Average-Neuro-Immune System -5 Fgure 8 Average approxmaton resdual usng sldng backpropagaton (ANN) vs Neuro-Immune system at zone (A, B, and C)

6 Also, the root mean squared errors (RMSE) of the approxmatons are calculated; n ( YActual YApp ) = RMSE( YApp ) = n (9) In (9), YActual and Y App are the actual and approxmated values whch are assgned to the nstantaneous temperature values of each under approxmaton node; n denotes the number of the approxmated records The RMSE of the average approxmaton resdual usng ANN and Neuro- Immune system s llustrated n Fgure 9 Consderng the Fgure 9, the Neuro-mmune system leads to more accurate approxmaton n comparson to the sldng backpropagaton 4 RMSE-ANN RMSE-Neuro-Immune System the arbtrary set ponts ncludng [ºC, ºC], [ºC, ºC], [-ºC, ºC], [ºC, ºC], and [-ºC, ºC]; the accuracy of the proposed Neuro-mmune system was hgher than sldng backpropagaton n all cases TABLE I COMPARISON OF DATA APPROXIMATION TECHNIQUES Data Approxmaton Accuracy (ºC) Average Calculaton Tme Sldng Backpropagaton ±5 Sec Neuro-mmune System ± 8 Sec The role of the classfcaton mechansm s to use the approxmaton values to classfy the records as plausble, mplausble and unknown To classfy the records, after each data approxmaton, the mmune system calculates the Eucldean dstance between the actual records (Ab) and approxmatons (Ag) at each nstance Fgure shows the nstantaneous calculated Eucldean functons n zone 5 A-Affnty Functon Fgure 9 RMSE of data approxmaton n zone The RMSE of the average approxmaton resdual for each cluster ncreases durng the frst mnutes of data approxmaton, and afterward remans mostly steady due to the nature of the RMSE functon Durng the test, the RMSE of the average approxmaton resdual n zone (ncludng nodes A, B, and C) reaches to about 45 and 9 (ºC) usng the sldng backpropagaton and Neuro-mmune systems, respectvely; the RMSE value for zone (ncludng nodes D, E, and F) reaches to 66 and 4 (ºC) usng the proposed algorthms respectvely, as seen n Fgure 4 RMSE-ANN RMSE-Neuro-Immune System Fgure RMSE of data approxmaton n zone Table I compares the accuracy of the neuro-mmune algorthm and sldng back propagaton The results show that despte ncreasng the calculaton tme for about 5 Sec, the Neuro-mmune data approxmaton s more accurate than ANN The proposed scenaro was tested n 5 dfferent envronmental condtons by adustng the reefer unt usng 5 B-Affnty Functon 4 C-Affnty Functon Fgure Eucldean dstance between sensor nodes Fgure llustrates that the Eucldean functon (EF) for each node remans less than 5 C when the records of all three sensor nodes are plausble Any devaton from ths range could be classfed as unknown or mplausble (whch was descrbed before as hgh accordng to Fgure 5) hen there s a fault n the wreless sensor network (lke defecton or battery dscharge), the records are assumed as mplausble whch causes a consderable devaton of data approxmaton from the actual records n the wreless sensor network Sometmes due to naccurate readngs or weakness n data approxmaton the data s classfed as unknown The unknown area has to be nvestgated whether the devaton of data approxmaton from the actual value s due to a fault occurrence n the wreless sensor network or not Three data classes are defned; Class : Plausble records (EF<5 C) Class : Unknown (plausble or mplausble) (5 C< EF< C) Class : mplausble records ( C< EF) To evaluate the unknown records (Class ), an absolute sldng correlaton functon s calculated to determne the relatonshps between sensor records [4]

7 ( s s )( s s ) ASC ( S, S ) = () ( s s ) ( s s ) ASC( S, S ) denotes the absolute sldng correlaton factor between S and S ; also, s and s are the elements of those two sets respectvely over the last four sequental records; s and s are ther average values Corr (A,C) (, 5, ) Corr (B,C) (,, 5) Z-Axs Z-Axs Y-Axs (a) (,, 5) (, 5, ) (b) Fgure Data classfcaton range Corr (A,B) Accordng to Fgure, when the absolute sldng correlaton factors between sensors are hgh, the unknown class s assumed as plausble (Class ); otherwse, when the EF s n unknown range and the sensor records are not hghly correlated, the records s classfed as mplausble (Class ) VI CONCLUSION In ths paper, an advanced bo-nspred data processng technque was presented for a wreless sensor network The mplemented technque conssts of two stages ncludng a Neuro-mmune data approxmaton and a data classfcaton mechansm Immune system was used to tran the neural network; the obtaned results showed that the developed Neuro-mmune approach leads to more accurate data approxmaton n comparson to the sldng backpropagaton technque To evaluate the relablty of the records n the wreless sensor network, the Eucldean dstances were calculated The future work s applcaton of the mmune system to optmze other network parameters such as number of neurons n order to mprove the data processng accuracy ACKNOLEDGMENT Ths research was supported by the German Research Foundaton (DFG), as part of the Collaboratve Research Centre 67 on "Autonomous Logstc Processes" REFERENCES [] N Mahalk, Sensor Networks and Confguraton: Fundamentals, Standards, Platforms, and Applcatons, Sprnger,USA, 7, pp [] R Verdone, D Dardar, G Mazzn, and A Cont, Sgnal processng and data fuson technques for SANs, reless Sensor and Actuator Networks, Academc Press, London, UK, 8, pp - 75 [] M M Gupta, L Jn, and N Homma, Statc and Dynamc Neural Networks: From Fundamentals to Advanced Theory, ley-ieee: NJ, USA, [4] A Jabbar, R Jedermann, R Muthuraman, and Lang, Applcaton of Neurocomputng for Data Approxmaton and Classfcaton n reless Sensor Networks, Sensors Journal (specal ssue on Neural Networks and Sensors), 9, Vol 9, pp [5] A Jabbar, R Jedermann, and Lang, Neural Network based Data Fuson n Food Transportaton System, th Internatonal Conference on Informaton Fuson, Cologne, Germany, 8, pp -8 [6] A Jabbar, R Jedermann, and Lang, Applcaton of Data Approxmaton and Classfcaton n Measurement Systems - Comparson of Neural Network and Least Squares Approxmaton, IEEE nternatonal conference on computatonal ntellgence for measurement systems and applcatons, Istanbul, Turkey, 8, pp [7] D Dasgupta, Artfcal Immune Systems and Ther Applcatons, Ed, Sprnger-Verlag, 999 [8] A Swam, Q Zhao, and Y Hong, reless Sensor Networks: Sgnal Processng and Communcatons Perspectves, John ley and Sons, 7, pp 65 [9] E-O Blass, J Horneber, and M Ztterbart, Analyzng Data Predcton n reless Sensor Networks, IEEE Vehcular Technology Conference, Sngapore, 8, pp [] A Heshmat and M Reza Soleyman, An Energy-Effcent Cooperatve Algorthm for Data Estmaton n reless Sensor Networks, Canadan Conference on Electrcal and Computer Engneerng, Canada, 7, pp 98 9 [] D Dasgupta, Artfcal Neural Networks and Artfcal Immune Systems: Smlartes and Dfferences, Proc of the IEEE Systems, Man and Cybernetcs, FL, USA, 997, pp [] L N de Castro, F J Von Zuben, and GA de Deus, The constructon of a Boolean compettve neural network usng deas from mmunology, Neurocomputng, vol 5,, pp 5 85 [] L N De Castro and F J Von Zuben, An Immunologcal Approach to Intalze Feedforward Neural Network eghts, Proc of Internatonal Conference on Artfcal Neural Networks and Genetc Algorthms, Prague, Czech Republc,, pp 6-9 [4] Crossbow Homepage Avalable onlne: last accessed date: 64 [5] R Past and L N De Castro, An Immune and a Gradent-Based Method to Tran Mult-Layer Perceptron Neural Networks, Internatonal Jont Conference on Neural Networks, Vancouver, Canada, July 6-, 6 4