Categorizing WSN's Sensory Data Using SelfOrganizing Maps

Abdul Mannari'and Haroon A. Babre'Department of Electronic and Computer Engineering, NFC Institute of Engg and Tech. Training, Multan, Pakistan.

2Department of Electrical Engineering, University of Engineering and Technology, Lahore, Pakistan.mannan@nfciet.edu.pk, babri@uet.edu.pk

Abstract-- Wireless Sensor Networks suffer greatly from theirlimited battery power whose utilization is increased manifolds asa node has to transmit or receive fairly large amount of data.Several algorithms, some for scheduling battery power e.g.Dynamic Voltage Scheduling, Static Voltage Scheduling,Dynamic Power Management etc and other with emphasis ondesigning efficient routing protocols have been designed in past.Some algorithms, however, address this very issue at softwarelevel by writing memory and CPU friendly programs. Thispaper proposes Self Organizing Maps (SOM) basedunsupervised Artificial Neural Network learning technique toenhance average battery life. Proposed system allows all activenodes to transmit their sensory data to the base station node(BSN) which has a 2X3 SOM running on it. Sensor nodes startsending data to the BSN; it keeps on making categories and putsrelevant data in appropriate categories/ classes. SOM is trainedafter it has received anum ber of such transmissions from activenodes. Class definitions are then broadcast to all active nodes byBSN and from then onwards they transmit only the classdefinitions (that are fairly lesser in size) to BSN and hencesignificant battery power is conserved. We have showed anoverall 48.5% battery power saving using the above technique.

Index-Term - Base Station Node, Self Organizing Maps, ArtificialNeural Networks, Wireless Sensor Networks, Sampling,Competition, Adaptation

I. INTRODUCTIONSelf Organizing Maps (SOM) is an unsupervised learningtechnique inspired by the physical arrangement of neuronswithin human brain. A Self Organizing Map can be one, twoor three dimensional depending on the complexity ofrecognition or prediction task and the computationalcapabilities of hardware [1].

A Wireless Sensor Network (WSN) consists of very low costand low power multifunctional sensors deployed in largenumber to monitor some physical activity happening withintheir surroundings e.g. movement of troops, measurement oftemperature, humidity, vibration etc. Their implementationbecame possible due to recent developments in micro-electro­mechanical systems (MEMS) and state of the artcommunication electronics. Many real world applications arepotential applications of Wireless Sensor Networks and havecaught the interest of researchers from variousinterdisciplinary fields due to their unique characteristics e.g.an environment with large number of sensors distributed atphysically scarce locations, flexibility of movement even

978-1-4244-4361-1/09/$25.00 ©2009 IEEE

after their deployment, self organization feature, rapiddeployment and inherent intelligence [2].SOMs have been applied to localize/ place sensor nodes inorder to reduce communication overheads [11]. ART andFuzzy ART algorithms have also been applied to sensornetworks to increase battery life time and tendency forhierarchy building [6]. On the other hand, some researchershave tried techniques like Static Voltage Scheduling andDynamic Power Management to activate electronic circuitryonly when it is required [12]. The battery's life timeenhancement is still an important problem. Efficient linklayer protocol design for the network has also been addressedwith interesting results [14]

This paper presents a novel scheme for implementing a twodimensional Self Organizing Map on a sample WirelessSensor Network to lower the volume of data packetstransmitted and received within the network and hencereduction in overall battery consumption is realized. Resultsshow that (1) Battery power saving monitored versus numberof data samples transmitted for training (2) Battery powersaving versus time interval between two consecutivetransmissions and finally (3) an optimum operating conditionwas calculated by combining both of the above twodimensional plots.

Section II discusses the details of Wireless Sensor Networks/IEEE 802.15.4 standard with possible routing architecture.Section III discusses Self Organization Maps briefly. SectionIV describes how sensory data is categorized and Section Vdescribes Results and Analyses

II. WIRELESS SENSOR NETWORKS

Recent developments in wireless communications, integratedcommunication electronic design and Micro ElectroMechanical Systems have led to the development of cheapand efficient Wireless Sensor Networks. They targetprimarily the very low cost and ultra low power consumptionapplications. Data throughput of the network and reliabilityof a node can however be compromised to be of secondaryconcern [2].

Non-critical and non-real time applications in residential,business and industrial sectors have accelerated the

Consumption

Cost Highest High Low

Size Large Small Very Small

requirement to implement inexpensive Wireless SensorNetworks and hence given rise to the concept of a modelcheap wireless personal area network (LR-WPANs: LowRate Wireless Personal Area Network) [2]. Consequently, inyear 2003, the LR-WPANs standard was accepted by IEEEand was given the name IEEE 802.15.4 [3]. Star topology Peer to Peer topology

Fig 1. Star and Peer to Peer topology

III. SELF ORGANIZING MAPS

2-127 bytes

o Active Node

Prea- Le- Payloadmble Start

ngth

1

4-bytes~.. •r- --1-- --I~"------~------..I

Fig 2. The] EEE 802.15.4 packet structure

The 802.15.4 standard uses a packet structure as in Fig. 2.Each packet or physical layer protocol data unit contains apreamble, a start of packet delimiter, a packet length andpayload field or physical layer service data unit. The payloadlength can vary from 2 to 127 bytes depending on size of databeing exchanged. In our design, preamble carries node IDwhile Payload bears data from various active nodes

Supposing input samples and the map weights to bemultidimensional real valued vectors, the three phases of thetraining algorithm are as follows:

An SOM [1] is a neural network that learns applicationinformation as a set of weights associated with the neurons.In comparison with other techniques (e.g. Multi-LayerPerceptron), SOMs are unique because the neurons arearranged in regular geometric structure which can be one ortwo dimensional. Three dimensional structures can also existbut are not very common. Neurons can be arranged either (1)Grid or (2) Hexagonal or (3) Random topology beforetraining starts.

Topology plays a central role in learning process and on finalarrangement of neurons after they have learned. Training ofSelf Organizing Maps in unsupervised i.e. total number oftarget classes is known before hand but their exact defmitionsare not obvious unless network is adequately trained.

@ Coordinator

1. Sampling:

802.11b 802.15.1 802.15.4

Range (m) 100 Up to 100 10

Throughput10 1 0.25 Max

(Mbps)Power High Low Very Low

IEEE 802.15.4 is distinct from other wireless technologies inthe sense that its goal is to achieve cheap short distancecommunication at low data rate and hence realization of anetwork of large number of very small sized nodes possiblein contrast with its competitors.

Wireless Sensor Network/ IEEE 802.15.4 are worth using intwo frequency bands (1) 868/915 MHz or (2) 2.4 GHz withthe same data packet structure. However data rate can rangefrom 20 kbps to 250 kbps as central carrier frequencyincreases from lowest value i.e. 868 MHz to maximum valuei.e. 2.4 GHz

The 802.15.4 LR-WPANs standard provisions two differenttopologies for network formation [3]:

• Star/ Hub Topology• Peer to Peer Topology

Inspired by conventional Ethernet Hub Architecture Startopology only allows all active nodes to be in 'directcommunication with Coordinator or Base station node. Twoactive nodes cannot be in direct communication. In Peer toPeer topology, in contrast, any node can communicate withany other node irrespective of active or base station node.Data communication in peer to peer topology is easier ascompared to star topology but at the cost of increasednetwork complexity.

TABLE ICOMPARISON OF VARIOUS WIRELESS ECHNOLOGIES

Suitability of Wireless Standards subsisted before the adventof IEEE 802.15.4 for applications insisting massivedeployment of Wireless Transceivers was dubious due totheir high cost and unnecessarily high data rate. IEEE802.15.4 however, has revolutionized the deployment ofvarious small sensor nodes in great numbers for monitoringvarious physical phenomenon, conveniently distributed atscarce locations and keeping their cost low at a compromiseof low data rate and range of communication. A comparisonof the 802.15.4 LR-WPANs with 802.11b WLAN and802.15.1 Bluetooth ™ is shown in Table 1.

An input sample taken from the training data set at a certainiteration is called x(n) and is applied to the network. Synapticweight vector corresponding to a neuron j is Wj' where jranges from 0 to t with t representing total number ofneurons.

WinningNeuron

Fig. 3: A 2-dimenstional Self-Organ izing Map

2. Competition:The sample x(n) is compared with the map weights of eachneuron through the use of a discriminating functionf=j{x,w).The neuron that scores the maximum value wins thecompetition and becomes the Winning Neuron (WN). Therecan be various choices for calculating distance associatedwith the discriminating function e.g. Euclidean, grid or box.However , if it is implemented using the Euclidean distance,the selection will be governed by the following expression:

3. Synaptic Adaptation:Finally, the synaptic weights associated with WinningNeuron along with its neighbors' are adapted according to thefollowing rule:

W j (n + 1)= wj (n) + .,,(n)h(j,WN(n))[x(n) - wj (n)]

Synaptic update is controlled by the global learning rateparameter IJ and by a neighborhood function h =h(i, j )Learning rate must decrease with increasing number ofiterations. Typical range of learning rate is from 0.1 to 0.01 asnumber of epochs increase. One possibl e choice for IJ is adecreasing exponential function given by:

lJ(n) = lJo exp(-n / k[)where no is initial value and k, is a time constant.

Synaptic update also involves neighborhood function whichdecides how much and how far neighboring neurons shouldbe affected in terms of their weights along with weight

updating of winning neuron . The function can be Gaussianand the effect should be decreasing as Euclidean distancefrom the winning neuron increases.

h( . .) (d(i,j)2)l ,j =exp -

2r(n)where d(i,j) is the distance between two neurons on the map.Weight change is significant near the winning neuron andbecomes negligible on the boundary of the neighborhoodrange. One last parameter to be discussed is r(n) which is theradial distance from the winning neuron and is calculated asmonotonously decreasing function:

r(n) =roexp(-n / k2 )

where ro is initial value and k2 is another time constant.

r(n) is high initially, ensures early convergence of thenetwork and becomes small as epochs increase, which causesindividual neurons to become sensitive for various featurescontained by input data set.

IV. CATEGORIZING SENSORY DATA

Wireless Sensor Networks suffer greatly due to inherent lossof battery power as they process information sensed fromtheir surroundings and then transmitting and! or receiving thisinformation. We have targeted this very drawback and havecome up with an unsupervised learning techniqu e i.e. SelfOrganizing Maps which is trained on sample two dimensionaldata collected from various active nodes. Once training iscomplete active nodes are allowed to transmit only the classtitles which is less bulky as compared to full sensed datatransmission. Sample input data was first plotted on a twodimensional scale to exploit any potential clusters present asshown in Fig. 4.

A Self Organizing Maps based network was trained with thesame training data. Output neurons were connected as a 2X3map as shown in Fig. 5.

A Self Organizing Map orders various data samplesaccording to their mutual similarity. The neighborhoodfunction used was a decreasing exponential function and avariable learning rate was selected to be 0.9 initially and thendecreased to O.Olas the last epoch reached.

Fig. 4: Sample two dimensional data

in1

in2

Fig. 5: A 2X3 Self-Organizing Map

TABLE IIPOWER CONSUMPTION OF VARIOUS OPERATIONS

OperationPowerConsumption

Central Processing Unit3mW

(CPU)Transmitter 30mWReceiver 40mWSleep 5JlW

There are two major factors affecting overall power savingsachieved by the scheme.

• Total number of training data samples• Interval between two consecutive transmissions

(after the network is trained)

inA

inB Jf/ ~: .-- --v:.;;::: .':::::.:~~:~.'.~~:~<,

'.<.'.

oute-~

outO-~

outE-~

outF-~

outG-~

outH-~

In each case the power saving was the calculated differencebetween "power consumed if there is no training" and "powerconsumed by training/ trained network".

If the total number of training data samples is kept low, thenetwork will not be able to converge. However, powerconsumption will be small. On the other hand, as the numberof training samples is increased, the network will beconverged adequately but at a cost of increased powerconsumption.

The time interval between two consecutive transmissions waskept constant and power saving was calculated for varyingvalues of training data samples which, in each case, werevaried from 18 to 23. This is done by reading data of Table 3one row at a time.

TABLE IIIPOWER SAVING CHART

Fig. 6: Feedforward network with 6 output neurons

The performance of the above 2X3 Self Organizing Map wascompared with a feed forward backpropagation network withone hidden layer and six output neurons as shown in Fig. 6.

The SOM was found advantageous in terms of its very earlyconvergence. Performance of Self Organizing Mapsalgorithm was verified for input samples described in nextsection.

v. RESULTS AND ANALYSES

Implementation of a 2x3 Self Organizing Map neural networklearning on one wireless sensor network active node canresult in a significant power saving which will be obvious aswe move further in this section.

Power dissipation of a typical WSN node for variousoperations is given in Table 2.

Interval Power Savingb/w (mW)consecutive Training Data Samplestransm- 18 19 20 21 22 23issions(hours)2 23 23 22.5 22.5 22 21.54 37 36 35 34.5 33.5 32.56 44.5 43 41.5 40.5 39 37.58 48.5 46.5 44.5 42.5 40.5 3910 49.5 47 44 42 39.5 37.512 48.5 45.5 42.5 39.5 37 34.514 44.5 41 38.5 34 30.5 27.516 38.5 34 30 26 22.5 1918 34.5 30.5 25.5 21 17 14.520 30 25.5 21 16.5 12.5 922 25.5 20 16.5 11.5 8 4.524 21 15.5 11 6.5 2 2

From 4 to 24 hrs interval between transmissions50r----,-----r----r-----,-------,

2520

-------~- - --------- ~ - - --"-,---"--

-- 08=18

-- 08=19

-- ---- -- 08=20

- - - - 08=21

- 08=22

~08=23

5

, ,, ,5 -- ---- - --- -~---~-'- ------- ~ - -- - - -- - - -- -~---- -- ----

, ,, ,, ,

10 ------ ---- -

50,----,-----::::;:;::=:::=----,---,------,

45

From 18 to 23 data samples

40

35zn

.!Oiii 30'"Q;:;: 25o

0-

g;, 20'"

?f. 15

232219

" ,, -..... ' , I-----------~ --------- .". :--.::: -.- - - - - - - - - ~ -----------~ -----------: : -- -- : :I - +-. ,

- - - - - - - -- --" ---- - -- - - ..----- --- - --.-----..:;-- - ---.------ - ----

-- TI=4 : i --' ~ -- _---.r --..--.-~ --.-.---.--:---.--_::-.... -..-- ·- TI=8 : ' :

-- ------- TI=12 - - - -~- - - - -- - - - -- ~ - - - - - - - - - ~ - - - - - - - - - - -

- ---- TI=16 ::- -- - j- - --- -- - - j - - - --- -- - -- i - - --- --

-----#--- TI=20 : : :---+- TI=24 --- -;--- ----- --- ;--- --- ---;--- --- -----

, ,, ,0'--------'-----'-----'------"-----'18

10

5

, ,~'"~,"~ I I :

:: : : : : : :::: ::; : : : : : : ~ :::: p~~;~:~:~; ~:~~: :::I I I - - " _ J.

_ _ ____~_ :-.. ~ _.. I --l-----------t:::::::::. .. .

Q)

OS: 25o

(L

~ 20

'"?ft. 15

'" 35.s~ 30

Fig. 7: Power Saving versus Training Data Samples Fig. 8: Power Saving versus Time Interval between transmissions

Six different plots were obtained by changing time intervalbetween two consecutive transmissions from 4 to 24 hours asshown in Fig 7.

A. POWER SAVING VS INTERVAL BETW EENRANSMISSIONS

After the network is trained, if transrmssions are madefrequent, the battery will exhaust drastically but with thebenefit of continuous monitoring. On the other hand, ifinterval between two consecutive transmissions is larger, thelife time of voltage source will increase (with someconditions discusses later in this section) but at a cost of weakmonitoring.

Training Data Samples were kept constant and power savingwas plotted versus interval between two transmissions whichin each case, was varied from 2 to 24 hours. This is done byreading data of Table 3 one column at a time.

50

40OJ.s>~ 30

~c0.. 20~

OJ~

"#. 10

o23

Number oftrainingdatasamples 18 0

········.L"" .....

........;

25

Timeinterval betwentransmissions (hrs)

Six different plots were created by varying trammg datasamples from 18 to 23 which are shown in Fig. 8 Fig 9. Power Saving versus Time Interval and Training Data Samples

Fig 8 reveals that the power saving increases initially withincreasing time interval between transmissions but as timeinterval crosses a certain value, it starts decreasing. It is dueto the fact that if data transmissions are not frequ ent then thepower consum ed to train the network initially, becomes anoverhead .

From this plot we can infer the optimum values of trainingdata samples and time interval between consecutivetransmissions required to maximize overall power saving.Optimum parameters derived from Fig 9 are given in Table 4.

TABLE IVOPTIMUM PARAMETERS

B. POWER SAVING VS TIME INTERVAL ANDTRAINING DATA SAMPLES

The effect of "Time interval between transmissions" and the"Number of training data samples" on "Power Saving" isshown in Fig 9.

Optimum number of training data samples 18Optimum Time interval between two 9.5 hrsconsecutive transmissionsMaximum Normalized Power Saving achieved 48.5%

VI. CONCLUSION

Wireless Sensor Networks suffer greatly due to inherent lossof battery power as they process information sensed fromtheir surroundings and then transmitting and! or receiving thisinformation. We have targeted this very drawback and havecome up with an unsupervised learning technique i.e. SelfOrganizing Maps which is trained on sample two dimensionaldata collected from various active nodes. Once training iscomplete active nodes are allowed to transmit only the classtitles which is less bulky as compared to full sensed datatransmission. We have shown that applying this technique totwo dimensional sensed data has resulted in enormous batterypower saving which is around 48.5%.

ACKNOWLEDGEMENT

The authors gratefully acknowledge Jameel Ahmed of NFCInstitute of Engineering and Technological Training, Multanand Asim Loan of University of Engineering andTechnology, Lahore for discussions and valuable input.

REFERENCES

[1] S. Haykin, Neural Network: 'A Comprehensive Foundation', 2nd

Edition, Prentice Hall, 1998[2] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci,

"A survey on sensor networks," IEEE CommunicationsMagazine, vol. 40, 2002, pp. 102-114

[3] G. 1. Pottie and W. 1. Kaiser, "Wireless integrated networksensors," Communications of the ACM, vol. 43, no. 5, 2000, pp.51-58

[4] S. Lindsey, K. M. Sivalingam, and C. S. Raghavendra, "Datagathering algorithms in sensor networks using energy metrics,"IEEE Transactions on Parallel and Distributed Systems, vol. 13,2002,pp.924-935

[5] P. Emmanauel , C. Nick, "Unsupervised categorization andcategory learning", The Quarterly Journal of experimentalpsychology, vol. 58, no. 4, 2005, pp-733-752

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

A. Kulakov, D. Davcev, "Tracking of unusual events in wirelesssensor networks based on artificial neural-networks algorithms",Proceedings ofthe ICIT, Vol. 2, pp-532-539, 2005Mohammad Ilyas and Imad Mahgoub, "Handbook of sensorNetworks: Compact Wireless and Wired Sensing Systems" CRCPress, 2005P. Radivojac, U. Korad, K. M. Sivalingam, Z. Obradovic,"Learning from Class-Imbalanced Data in Wireless SensorNetworks", Vehicular Technology Conference, Vol. 5, 2003,pp.3030-3034Ethem Alpaydin, "Introduction to Machine Learning", The MITPress, 2004T. Kohonen, "Self-organized formation of topologically correctfeature maps" Biological Cybernetics, Vol. 43, no. 1, 1982, pp.59-69Gianni Giorgetti, Sandeep K. S. Gupta, Gianfranco Manes,"Wireless Localization Using Self-Organizing Maps",Proceedings of the 6th International conference on Informationprocessing in sensor networks, 2007, pp. 293-302Min, R. Bhardwaj, M. Seong-Hwan Cho Shih, E. Sinha, A.Wang, A. Chandrakasan, A. "Low-power wireless sensornetworks", Proceedings of the Fourth International Conferenceon VLSI Design, 2001, pp. 205-210Vijay Degalahal, Tim Tuan, "Methodology for High LevelEstimation of FPGA Power Consumption", Proceedings of the2005 conference on Asia South Pacific design automation, 2005,pp. 657 - 660, 2005Tan et aI, "Power efficient data gathering and aggregation inwireless sensor networks", SIGMOD Record, Vol. 32, No.4,December 2003A. Mannan "Realization of an unsupervised learning scheme forcategorization of sensory data of WSNs" MSc Thesis,Department of Electrical Engineering, UET, Lahore, Pakistan Oct2007