Professional Documents
Culture Documents
1
means structural condition is healthy and H means
I unhealthy. So, the power set of X has four elements:
Be/(¢)= 0,
L m(B) · (4)
X
Be/(A)=
J
- -
. B
'e null set, hI = {H}, hz = {H} , and h3 = = {H ,H} .
Suppose that the basic belief (reliability) of sensor Sj is
The Bel(A) represents the weight of evidence supporting PI. If node Sj provides the measurement that is healthy,
A's provability, On the other hand, the upper probability then its mass value assignment (according Dempster
function, or plausibility function, PI, measures how Shafer) will be the following: m;Chl) Pj' m;Chz) 0, and
= =
much the information given by source does not m;Ch3) I - Pj' If node Sk provides the measurement that
=
contradict a specified element as the right answer. In is unhealthy, its basic probability assignment will be the
other words, the plausibility is the weight of following: mihl) =0, mk(hz) P k, and mk(h3) I - Pk. In
= =
evidence that do not refute B. The plausibility function this way, a cluster-head can have the mass values based
to the element A is: on the decisions of its n subordinate sensors. Then,
according to Dempster's rule of combination in (6), the
1
I
PI(¢)=O,
combined mass values can be calculated using
PI(A)= . L m(B) · (5)
�nH*¢ J 1
I
m(¢)=O,
"
There are several ways of taking the final decision. For L TI m (S )
, , i, (8)
.
instance, decision can be made by choosing the meA)= v,,,,,,��,�
hypothesis with the maximum mass, belief or J
plausibility.
where K represents the degree of conflict given by
B. 1 Rules of combination of evidences from
different observers: Evidences obtained from two
independent sources and expressed by the two mass (9)
functions ml and mz on the same power set can be
properly combined to obtain a joint mass function ml,z,
The Dempster's rule of combination is expressed by Thus the combined mass values, from the evidences
orthogonal sum and normalization as given by provided by the n sensors, can be calculated for hI and
hz. The fusion sensor finally prepares the report based
m(¢)=O,
on the higher value between m(hl) and m (hz).
p
VI. CONCLUSIONS
and �"" be the immediate past and current,
In this paper, we have presented a protocol for structural
respectively, reliabilities of sensor Sj at the kth decision health monitoring using wireless sensor networks. Here
interval. The updated reliability p � of Sj' to be used every cluster-head uses the Dempster-Shafer theory of
during the (k+ l)th decision interval, is calculated using evidence to take the final decision while aggregating
different measurements for the subordinate nodes in that
cluster. Results through simulations show that the
(10) proposed method achieves higher reliabilities than the
existing event detection mechanisms.
where t5 is the tuning parameter that controls the rate of
update, and p �u" is calculated by REFERENCES
ClIrr
[1) F.-K. Chang, Structural Health Monitoring: The
L
PJ (11)
:::: _
W
1 ' Demands and Challenges. CRC Press, 2001.
[2] J. P. Lynch, A. Sundararajan, K. H. Law, A. S.
Kiremidjian, and E. Carryer, "Power-efficient data
where W is the window size representing the number of
past readings to consider and f is the number of management for a wireless structural monitoring
incorrect readings in this window. system, " Proc. of the 4th International Workshop
on Structural Health Monitoring, 2003, pp. 15-17.
V. PERFORMANCE EVALUATION [3] N. Xu et al., "A wireless sensor network for
structural monitoring, " Proc. of SENSYS. ACM
In this section, we evaluate the performance of the Press, 2004, pp. 13-24.
proposed fault-tolerant data collection protocol through
[4] K. Mechitov, W. Kim, G. Agha, and T. Nagayama,
simulations using C++. We examine the effectiveness of
"High-frequency distributed sensing for structure
the proposed protocol in terms of the percentage of
monitoring, " Proc. of First IntI. Workshop on
detection errors in the reports generated by the cluster
Networked Sensing Systems (INSS), 2004.
heads.
To compare the detection performance of the proposed [5] J. P. Lynch, A. Sundararajan, K. H. Law, A. S.
Dempster-Shafer theory-based detection method, we Kiremidjian, E. Carryer, H. Sohn, and C. R. Farrar,
implement two well-known methods such as Bayesian "Field validation of a wireless structural monitoring
detection and Neyman-Pearson detection method [11]. system on the alamosa canyon bridge, " Proc. of
We also applied the method based on k-out-of-n rule SPIE's 10th Annual International Symposium on
presented in [8], which attempts to minimize the Smart Structures and Materials, San Diego, CA,
detection error in the Bayesian and Neyman-Pearson USA, March 2-6, 2003.
detections. [6] B. Krishnamachari, and S. Iyengar, "Distributed
TABLE!
bayesian algorithms for fault-tolerant event region
AVERAGE DETECTION ERRORS (%) MADE BY CLUSTER-HEADS WHEN detection in wireless sensor networks, " IEEE
APPLIED DIFFERENT METHODS Transactions on Computers, vol. 53, 2004, pp.
Ba esian Neyman-Pearson Proposed 241-250.
n Original k-out-of-n Original k-out-of-n method
[7] D. Koller, and N. Friedman, Probabilistic
3 20.49 1 2.43 20. 1 0 1 3. 1 1 9.02
5
Graphical Models - Principles and Techniques.
21.10 8.30 20.33 8.86 7.76
7 2 1 .45 6.0 1 20.35 6.43 5.23 MIT Press, 2009.
9 2 1 .60 4.46 20. 1 5 5.65 3.8 1 [8] X. Luo, S. Member, M. Dong, and Y. Huang, "On
n number of sensors III each cluster. distributed faulttolerant detection in wireless sensor
networks, " IEEE Transactions on Neural Networks,
=
We test the proposed detection scheme in several vol. 55, 2006, pp. 58-70.
simulated sensor networks. We use 100 clusters in our [9] S. Salicone, Measurement Uncertainty: An
simulations. We vary number of the subordinate sensor Approach via the Mathematical Theory of
nodes under the supervision of each cluster-head. Evidence. Springer, 2006.
Therefore, the total number of sensors is different in
[10]1. R. Boston, "A signal detection system based on
different simulations. Every result presented in this
dempster-shafer theory and comparison to fuzzy
detection, " IEEE Transactions on Systems, Man,
section is the calculated average of 1000 runs of a
simulation. The summary of results of our simulations
and Cybernetics, vol. 30, pp. 45-51.
is presented in Table 1. The probability of sensor faults
is kept 10% for all the simulations. [11] P.K. Varshney, Distributed Detection and Data
Table 1 clearly demonstrates that the error rates in the Fusion. New York: John Wiley & Sons, 1991.
structural health report done by the cluster-heads are