2. requirement to implement inexpensive Wireless Sensor Consumption
Networks and hence given rise to the concept of a model
cheap wireless personal area network (LR-WPANs: Low Cost Highest High Low
Rate Wireless Personal Area Network) [2]. Consequently, in Size Large Small Very Small
year 2003, the LR-WPANs standard was accepted by IEEE
and was given the name IEEE 802.15.4 [3]. Star topology Peer to Peer topology
Suitability of Wireless Standards subsisted before the advent
of IEEE 802.15.4 for applications insisting massive
deployment of Wireless Transceivers was dubious due to
their high cost and unnecessarily high data rate. IEEE
802.15.4 however, has revolutionized the deployment of
various small sensor nodes in great numbers for monitoring
various physical phenomenon, conveniently distributed at @ Coordinator o Active Node
scarce locations and keeping their cost low at a compromise
of low data rate and range of communication. A comparison
Fig 1. Star and Peer to Peer topology
of the 802.15.4 LR-WPANs with 802.11b WLAN and
802.15.1 Bluetooth ™ is shown in Table 1. ~
1
..4-bytes •r- --1-- 2-127 bytes
--I~"------~------..I
IEEE 802.15.4 is distinct from other wireless technologies in
the sense that its goal is to achieve cheap short distance Prea- Le-
Start Payload
communication at low data rate and hence realization of a mble ngth
network of large number of very small sized nodes possible
in contrast with its competitors. Fig 2. The] EEE 802.15.4 packet structure
The 802.15.4 LR-WPANs standard provisions two different The 802.15.4 standard uses a packet structure as in Fig. 2.
topologies for network formation [3]: Each packet or physical layer protocol data unit contains a
• Star/ Hub Topology preamble, a start of packet delimiter, a packet length and
payload field or physical layer service data unit. The payload
• Peer to Peer Topology
length can vary from 2 to 127 bytes depending on size of data
Inspired by conventional Ethernet Hub Architecture Star
topology only allows all active nodes to be in 'direct being exchanged. In our design, preamble carries node ID
while Payload bears data from various active nodes
communication with Coordinator or Base station node. Two
active nodes cannot be in direct communication. In Peer to
III. SELF ORGANIZING MAPS
Peer topology, in contrast, any node can communicate with
any other node irrespective of active or base station node.
An SOM [1] is a neural network that learns application
Data communication in peer to peer topology is easier as
information as a set of weights associated with the neurons.
compared to star topology but at the cost of increased
In comparison with other techniques (e.g. Multi-Layer
network complexity.
Perceptron), SOMs are unique because the neurons are
arranged in regular geometric structure which can be one or
Wireless Sensor Network/ IEEE 802.15.4 are worth using in
two dimensional. Three dimensional structures can also exist
two frequency bands (1) 868/915 MHz or (2) 2.4 GHz with
but are not very common. Neurons can be arranged either (1)
the same data packet structure. However data rate can range
Grid or (2) Hexagonal or (3) Random topology before
from 20 kbps to 250 kbps as central carrier frequency
training starts.
increases from lowest value i.e. 868 MHz to maximum value
i.e. 2.4 GHz
Topology plays a central role in learning process and on final
TABLE I arrangement of neurons after they have learned. Training of
COMPARISON OF VARIOUS WIRELESS ECHNOLOGIES Self Organizing Maps in unsupervised i.e. total number of
target classes is known before hand but their exact defmitions
are not obvious unless network is adequately trained.
802.11b 802.15.1 802.15.4
Supposing input samples and the map weights to be
Range (m) 100 Up to 100 10
multidimensional real valued vectors, the three phases of the
Throughput training algorithm are as follows:
10 1 0.25 Max
(Mbps)
Power High Low Very Low 1. Sampling:
3. An input sample taken from the training data set at a certain updating of winning neuron . The function can be Gaussian
iteration is called x(n) and is applied to the network. Synaptic and the effect should be decreasing as Euclidean distance
weight vector corresponding to a neuron j is Wj' where j from the winning neuron increases.
ranges from 0 to t with t representing total number of
. .) (d(i,j)2)
neurons. h( l , j = exp -
2r(n)
Winning
Ne uron where d(i,j) is the distance between two neurons on the map.
Weight change is significant near the winning neuron and
becomes negligible on the boundary of the neighborhood
range. One last parameter to be discussed is r(n) which is the
radial distance from the winning neuron and is calculated as
monotonously decreasing function:
r(n) = ro exp(-n / k 2 )
where ro is initial value and k2 is another time constant.
r(n) is high initially, ensures early convergence of the
network and becomes small as epochs increase, which causes
individual neurons to become sensitive for various features
contained by input data set.
Fig. 3: A 2-dimenstional Self-Organ izing Map
IV. CATEGORIZING SENSORY DATA
2. Competition: Wireless Sensor Networks suffer greatly due to inherent loss
The sample x(n) is compared with the map weights of each of battery power as they process information sensed from
neuron through the use of a discriminating functionf=j{x,w). their surroundings and then transmitting and! or receiving this
The neuron that scores the maximum value wins the information. We have targeted this very drawback and have
competition and becomes the Winning Neuron (WN) . There come up with an unsupervised learning techniqu e i.e. Self
can be various choices for calculating distance associated Organizing Maps which is trained on sample two dimensional
with the discriminating function e.g. Euclidean, grid or box. data collected from various active nodes. Once training is
However , if it is implemented using the Euclidean distance, complete active nodes are allowed to transmit only the class
the selection will be governed by the following expression: titles which is less bulky as compared to full sensed data
transmission. Sample input data was first plotted on a two
dimensional scale to exploit any potential clusters present as
shown in Fig. 4.
3. Synaptic Adaptation: A Self Organizing Maps based network was trained with the
Finally, the synaptic weights associated with Winning same training data. Output neurons were connected as a 2X3
Neuron along with its neighbors' are adapted according to the map as shown in Fig. 5.
following rule:
A Self Organizing Map orders various data samples
Wj (n + 1) = w j (n) + .,,(n)h(j, WN(n))[x(n) - w j (n)] according to their mutual similarity. The neighborhood
function used was a decreasing exponential function and a
Synaptic update is controlled by the global learning rate variable learning rate was selected to be 0.9 initially and then
parameter IJ and by a neighborhood function h =h(i, j ) decreased to O.Olas the last epoch reached.
Learning rate must decrease with increasing number of
iterations. Typical range of learning rate is from 0.1 to 0.01 as
number of epochs increase. One possibl e choice for IJ is a
decreasing 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 which
decides how much and how far neighboring neurons should
be affected in terms of their weights along with weight
4. Fig. 4: Sample two dimensional data
TABLE II
in1 POWER CONSUMPTION OF VARIOUS OPERATIONS
Power
in2 Operation
Consumption
Central Processing Unit
3mW
(CPU)
Transmitter 30mW
Receiver 40mW
Sleep 5JlW
There are two major factors affecting overall power savings
achieved by the scheme.
• Total number of training data samples
Fig. 5: A 2X3 Self-Organizing Map
• Interval between two consecutive transmissions
(after the network is trained)
oute
-~
In each case the power saving was the calculated difference
between "power consumed if there is no training" and "power
outO consumed by training/ trained network".
-~
If the total number of training data samples is kept low, the
inA
outE network will not be able to converge. However, power
-~
consumption will be small. On the other hand, as the number
of training samples is increased, the network will be
outF converged adequately but at a cost of increased power
-- --v:.;;::: ~: . .
Jf/ -~
consumption.
inB
outG
The time interval between two consecutive transmissions was
':::::.:~~:~.'.~~:~
<,
-~ kept constant and power saving was calculated for varying
values of training data samples which, in each case, were
'. varied from 18 to 23. This is done by reading data of Table 3
<.
'. outH
-~ one row at a time.
TABLE III
POWER SAVING CHART
Fig. 6: Feedforward network with 6 output neurons
Interval Power Saving
The performance of the above 2X3 Self Organizing Map was b/w (mW)
compared with a feed forward backpropagation network with consecutive
one hidden layer and six output neurons as shown in Fig. 6. Training Data Samples
transm-
18 19 20 21 22 23
issions
The SOM was found advantageous in terms of its very early (hours)
convergence. Performance of Self Organizing Maps
2 23 23 22.5 22.5 22 21.5
algorithm was verified for input samples described in next
4 37 36 35 34.5 33.5 32.5
section.
6 44.5 43 41.5 40.5 39 37.5
v. RESULTS AND ANALYSES 8 48.5 46.5 44.5 42.5 40.5 39
10 49.5 47 44 42 39.5 37.5
Implementation of a 2x3 Self Organizing Map neural network 12 48.5 45.5 42.5 39.5 37 34.5
learning on one wireless sensor network active node can 14 44.5 41 38.5 34 30.5 27.5
result in a significant power saving which will be obvious as 16 38.5 34 30 26 22.5 19
we move further in this section. 18 34.5 30.5 25.5 21 17 14.5
20 30 25.5 21 16.5 12.5 9
Power dissipation of a typical WSN node for various 22 25.5 20 16.5 11.5 8 4.5
operations is given in Table 2. 24 21 15.5 11 6.5 2 2
5. From 4 to 24 hrs interv between transmissions
al
50 r - - - - , - - - - - r - - - - r - - - - - , - - - - - - - ,
, ,
From 18 to 23 data samples
~'"~," ~
5 0 , - - --,-----:::;:;::=:::= - -- - , - - - , - - - - - - ,
:
:: :::::::::::;::::::~::::p~~;~:~:~;~:~~::::
I I :
45
I I I - - " _ J. 40
'" 35
_ _ ____ ~_ :-.. ~_ .. I --l-----------t:::::::::. . .
.s zn
35
" ,
.!O
~ 30 -----------~ --------- .". :--.::: -.- - - - - - - - - ~ ----------- ~ -----------
, -..... ' , I
Q)
: :
I
-- -- :
- +-.
:
,
..----- --- - --.-----..:;-- - ---.------ - ----
iii 30
'"
Q;
- - 08=18
OS:
o
25 - - - - - - - -- --" ---- - -- - -
:;: 25
---.r --..--.-i--.-.---.--:---.--_-. . -..-
--' ~ -- _ :
(L
- - TI=4 :
o - - 08=19
0-
~ 20 ~
g;, 20 -- ---- -- 08=20
- · - TI=8 : ' :
'"
?ft. 15 -- ------- TI=12 - - - -~- - - - -- - - - -- ~ - - - ------~-----------
'" - - - - 08=21
?f. 15
- ---- TI=16 : : - 08=22
10 - -- - j- - --- -- - - j - - - --- -- - -- i - - --- -- 10 - ----- ---- - -------~- - --------- ~- - --"-,---"--
TI=20 : : : ,
, ,
,
-----#--- ~ 0 8 =23
5 ---+- TI=24 --- -;--- ----- --- ;--- --- ---;--- --- ----- 5 -- ---- - --- - ~---~-'- ------- ~ - -- - - -- - - -- -~ ---- -- ----
, , ,
, ,
,
, , , ,
0'--------'-----'-----'------"-----'
18 19 22 23 5 20 25
Fig. 7: Power Saving versus Training Data Sampl es Fig. 8: Power Saving versus Time Interval between tran smissions
Six different plots were obtained by changing time interval
between two consecutive transmissions from 4 to 24 hours as
shown in Fig 7.
········.L
50 "" .
....
A. POWER SAVING VS INTERVAL BETW EEN
RANSMISSIONS ;
........
40
OJ
.s
After the network is trained, if transrmssions are made >
~ 30
frequent, the battery will exhaust drastically but with the
benefit of continuous monitoring. On the other hand, if
~
c
0.. 20
interval between two consecutive transmissions is larger, the ~
OJ
~
life time of voltage source will increas e (with some "#. 10
conditions discusses later in this section) but at a cost of weak
monitoring. o
23
Training Data Samples were kept constant and power saving 25
was plotted versus interval between two transmissions which
in each case, was varied from 2 to 24 hours. This is done by
reading data of Table 3 one column at a time. 18 0
Number oftraining data samples Timeint rval betwentrans ssions (hrs)
e mi
Six different plots were created by varying trammg data
samples 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 with From this plot we can infer the optimum values of training
increasing time interval between transmissions but as time data samples and time interval between consecutive
interval crosses a certain value, it starts decreasing. It is due transmissions required to maximize overall power saving.
to the fact that if data transmi ssion s are not frequ ent then the Optimum parameters derived from Fig 9 are given in Table 4.
power consum ed to train the network initially, becom es an
TABLE IV
overh ead . OPTIMUM PARAMETERS
B. POWER SAVING VS TIME INTERVAL AND Optimum number of training data samples 18
TRAINING DATA SAMPLES Optimum Time interval between two 9.5 hrs
consecutive transmissions
The effect of "Time interval between transmissions" and the
Maximum Normalized Power Saving achi eved 48.5%
"N umber of training data samples" on " Power Saving" is
shown in Fig 9.
6. VI. CONCLUSION
Wireless Sensor Networks suffer greatly due to inherent loss [6] A. Kulakov, D. Davcev, "Tracking of unusual events in wireless
of battery power as they process information sensed from sensor networks based on artificial neural-networks algorithms",
their surroundings and then transmitting and! or receiving this Proceedings ofthe ICIT, Vol. 2, pp-532-539, 2005
[7] Mohammad Ilyas and Imad Mahgoub, "Handbook of sensor
information. We have targeted this very drawback and have
Networks: Compact Wireless and Wired Sensing Systems" CRC
come up with an unsupervised learning technique i.e. Self Press, 2005
Organizing Maps which is trained on sample two dimensional [8] P. Radivojac, U. Korad, K. M. Sivalingam, Z. Obradovic,
data collected from various active nodes. Once training is "Learning from Class-Imbalanced Data in Wireless Sensor
Networks", Vehicular Technology Conference, Vol. 5, 2003,
complete active nodes are allowed to transmit only the class
pp.3030-3034
titles which is less bulky as compared to full sensed data [9] Ethem Alpaydin, "Introduction to Machine Learning", The MIT
transmission. We have shown that applying this technique to Press, 2004
two dimensional sensed data has resulted in enormous battery [10] T. Kohonen, "Self-organized formation of topologically correct
feature maps" Biological Cybernetics, Vol. 43, no. 1, 1982, pp.
power saving which is around 48.5%.
59-69
[11] Gianni Giorgetti, Sandeep K. S. Gupta, Gianfranco Manes,
ACKNOWLEDGEMENT "Wireless Localization Using Self-Organizing Maps",
Proceedings of the 6th International conference on Information
processing in sensor networks, 2007, pp. 293-302
The authors gratefully acknowledge Jameel Ahmed of NFC
[12] Min, R. Bhardwaj, M. Seong-Hwan Cho Shih, E. Sinha, A.
Institute of Engineering and Technological Training, Multan Wang, A. Chandrakasan, A. "Low-power wireless sensor
and Asim Loan of University of Engineering and networks", Proceedings of the Fourth International Conference
Technology, Lahore for discussions and valuable input. on VLSI Design, 2001, pp. 205-210
[13] Vijay Degalahal, Tim Tuan, "Methodology for High Level
Estimation of FPGA Power Consumption", Proceedings of the
REFERENCES 2005 conference on Asia South Pacific design automation, 2005,
pp. 657 - 660, 2005
[1] S. Haykin, Neural Network: 'A Comprehensive Foundation', 2nd [14] Tan et aI, "Power efficient data gathering and aggregation in
Edition, Prentice Hall, 1998 wireless sensor networks", SIGMOD Record, Vol. 32, No.4,
[2] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, December 2003
"A survey on sensor networks," IEEE Communications [15] A. Mannan "Realization of an unsupervised learning scheme for
Magazine, vol. 40, 2002, pp. 102-114 categorization of sensory data of WSNs" MSc Thesis,
[3] G. 1. Pottie and W. 1. Kaiser, "Wireless integrated network Department of Electrical Engineering, UET, Lahore, Pakistan Oct
sensors," Communications of the ACM, vol. 43, no. 5, 2000, pp. 2007
51-58
[4] S. Lindsey, K. M. Sivalingam, and C. S. Raghavendra, "Data
gathering 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 and
category learning", The Quarterly Journal of experimental
psychology, vol. 58, no. 4, 2005, pp-733-752