Professional Documents
Culture Documents
Accumulation Point Model: Towards Robust and Reliable Deployment
Accumulation Point Model: Towards Robust and Reliable Deployment
maximize the lifetime of sensor network. Most of them were intersected barrier belts.
for area coverage, such as [6] and [8]. Only two algorithms
3.4 Redundancy Elimination and Minimum Barrier
are proposed for barrier coverage [13]. It uses IBM and
Belts
RPM to address the sleep-wakeup problem for homogeneous
lifetime and heterogeneous lifetime. A barrier graph contains a set of sensors that are active for
Another important metric of sensor network in barrier barrier coverage. But what is more meaningful is a
coverage is the fault tolerance. Many papers propose the use minimum set of sensors to meet the QoS. The redundant
high sensor density to provide redundancy [1][5][10][15]. It sensors need to be eliminated or scheduled to sleep.
is generally believed that sensors are relatively economical
and unreliable devices, therefore high density is required to
provide a reasonable level of fault tolerance. This high
density assumption belongs to RPM for both area coverage
and barrier coverage. To change a coverage graph to a barrier graph, first we have
For barrier coverage, the centralized algorithms are to eliminate the round trip edges. We call this kind of
introduced in [11][13]. An L-zone localized algorithm is elimination to be Type I Elimination. There is a very special
provided in [2] for k-barrier coverage. Our APM uses a kind of redundant node - the single edge node. If a node (not
localized algorithm based on accumulation points, a subset including the virtual terminal nodes) has only one edge in a
of sensor nodes that locate closely together. coverage graph, then it has to be eliminated in a barrier
Weak barrier coverage and strong barrier coverage were graph. This is called Type II Elimination. The result of type
introduced in [11]. Weak barrier coverage assumes only I, II eliminations is a barrier belt. However, this barrier belt
when the intruder does not know the traverse path will it be is not the minimum barrier belt. So we need Type III
detected with high probability. The QoS improvements are Elimination, by which we eliminate the redundant nodes, so
discussed in [3][12]. In this paper, we assume the quality is that we can use the minimum number of nodes to meet QoS
important and all of the assumptions are based on strong requirement. We get two reasons for type III elimination:
barrier coverage. first, we can schedule the redundant nodes to sleep, so that if
an active node depletes, a redundant node can take its place;
3. Preliminary and Network Model second, the redundant nodes are also possible to be utilized
by other barrier belts.
3.3 Barrier Belt and Barrier Graph
Assume the coverage of sensor is a unit disk with sensing
radius R. A Coverage Graph G = (V,E) indicates that for
the active sensor set V , for ∀u, v V , if the distance duv
<2R, then there is an edge euv E. There are two virtual
nodes s, t V, corresponding to the left and right
boundaries. Barrier Space is the field of interest for barrier
coverage. It is a long strip physical region with arbitrary
shape and deployed by the barrier sensors.
From above introduction, we can summarize the Let us take Fig. 1 as an example, to show more details about
properties of the barrier belt as three types of elimination. First we get a coverage graph
1) There is no loop in a barrier belt; from the coverage relationship of sensors. The edge set {eaf,
2) It is possible that there are crossing edges of two efb, ecd} ⊂ V . But as for a barrier graph, every node can
distinct barrier belts in a barrier graph; only has two edges to support the barrier. So in this barrier
3) Except for the virtual terminal nodes {s, t}, there is no graph, edges {eaf , efb, ecd} are eliminated. However, this is
shared node for two distinct barrier belts in a barrier graph. only one barrier graph. There are several other barrier
graphs for this example; this leaves as an exercise for
Theorem 1. A sensor network barrier space is k-barrier readers. For this barrier graph Gb = (Vb,Eb). Since {eab, ecd}
coverage iff ∃k barrier belts in the barrier graph. are edges of the coverage graph, e and f are redundant
Proof: If each crossing path is covered by k distinct sensors, nodes. After the elimination of e, f, we get the final
then it must be that these k distinct sensors belong to k not minimum barrier graph. If a node is eliminated from the
minimum barrier graph, and it is not utilized by other
(IJCNS) International Journal of Computer and Network Security, 25
Vol. 2, No. 10, 2010
So we get lr ≤ 2Rnr/t2.
If the length lr = l is given, then we know that the number
of sensors required to support k-barrier coverage under RPM
is
26 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 10, 2010
Observation 1. For a barrier coverage system in a barrier have r = min(Rθ, R(1 − 2θ)/2). Let Rθ ≥ R(1 − 2θ)/2,we
know the permitted deploy region is a circle with radius r,
space, if it is constructed by n Akt accumulation points, then
then we can get Equ. (6).
the activation rate ρa = k/t.
Definition 6. Accumulation Deduction An accumulation 5. Comparison of the Barrier models
deduction is a decomposition and combination of an
accumulation degree, to make a conversion of a high 5.1 Connection Degree
accumulation degree to lower accumulation degrees and The connection degree for a sensor node A is the number of
vice versa, denoted as active nodes within the communication distance of A. Here,
the node we are referring is a normal node, not including
Let us look at an example: A35 = A23 + A12 = 3 A11 + 2 A10 . the virtual terminal nodes. We assume the communication
distance Rc ≥2R.
Here we know for a 3-barrier coverage system, we can use
3 Observation 2. For barrier coverage in k A11 + ( t − k ) A10
A5 APM in one barrier space. If we put them into two
independent belt model, the connection degree is at least 2.
barrier spaces, we can use A23 and A12 APM. We can also
Observation 3. For barrier coverage in Akt accumulation
use the IBM 3 A11 + 2 A10 . Here we have 3 A11 working belts, point model, the connection degree is at least 3k − 1.
and 2 A10 backup belts. As for each internal node A, it has two sets of neighbors,
each set has k active nodes. Also A has k – 1 brothers.
4.3.3 Analysis of Barrier Length
The barrier length of IBM and APM follows the same Observation 4. For barrier coverage in Rtk random point
schema. model, the connection degree is at least πk2/2.
Theorem 5. For a barrier coverage deployed in Akt Each node lies in a square (Equ. (2)). There are at least t
nodes in each cross path, k of them active. The nodes with
accumulation point model, the total number of nodes na =
least connection degree are the border nodes. We have
t・n, the overlap factor θ, then the barrier coverage length
(2R)2/2 (2R)2 = n/k2. So, we get n = πk2/2.
la = 2R(1−θ)n.
5.2 Network Lifetime
If the length is given, la = l, then the number of nodes na
required is If we do not concern the different roles of sensors, we can
assume all sensors’ lifetime is independent and follow the
same distribution. Considering a sensor’s lifetime is mainly
determined by battery, we can use the normal distribution to
From Equ. (3), it is easy to know that under the same
interpret the lifetime of a sensor node, with random variable
activation rate ρa = k/t, the same barrier length, the cost of
X to indicate a sensor node’s lifetime, the mean lifetime as
the RPM compared with the APM is
μ and variance σ, X ∼ N(μ, σ2).
The cumulative distribution function (cdf) is as follows:
Assume there are n accumulation points in the belt. Denote probability of a node’s lifetime defined in Equ. (7). Then
random variable Z as the lifetime of the belt. Then Fap,z(z)
={ the probability that Z ≤ z} = P(Z ≤ z). Denote py as the
probability of an accumulation point’s lifetime. Then
accumulation point, we know that the neighbors and else {after a period pw, A cannot receive k messages
brothers of A can receive the broadcasting message. After a back}
node B received A’s “I am here” message, B can compute A invokes the barrier building process;
the distance dAB, end if
if dAB < R then In the above algorithm, we use a wake-up query scheme
A is B’s brother; instead of an activate scheme. The activate scheme is
else {dAB < 2R} unpractical considering the current technique [8]; if a sensor
A is B’s neighbor; went to sleep, we cannot wake it up unless the sensor wake
else up itself by a short period. In the wake up scheme, it is
discard this message; possible that during the sleep period of a node, there is a
end if brother node depleted. And so it is possible to exist an
After this process, each sensor will construct a list of interval that the belt degrades to (k − 1)-barrier coverage.
Let’s denote the period from a node begin to report
brothers. Therefore, a sensor can learn the value of t in Akt
depleting to the time it totally depletes as pd. If we set the
by itself. After the node find out the value of t, it will fixed sleep period ps = 2pd, and the random sleep delay dr
broadcast a message “my accumulation point”, including the [0, 2pd], then even in the worst case, that is, t degrades to
list of its brothers. Usually, a neighbor node can receive the k+1, only one brother node left as a backup node, the
“my accumulation point” message 2t times from its maximum possible degraded monitoring interval max(ps +
neighbors. It simply separates the neighbors to be the left dr − 2pd) = 2pd.
neighbors and right neighbors. This finishes our neighbor If we decrease ps, the maximum possible degraded
finding process. monitoring interval max(ps + dr − 2pd) can be decreased. For
When the system begins the building process, a node can k > 1, the wake-up time is not a problem for APM, since
be chosen randomly to be the Seed. The seed is informed the node switch in each accumulation point is done locally, even
value of k in Akt . The seed selects k −1 brothers from its if one belt degrades a little time due to node switch, the
other belts are still working. While in IBM, unless we
brother list to be the Heads of the barrier belts. After a node
assume nodes can be activated, or we use the wake-up
A is appointed to be the head, it backoffs a random time,
scheme and set dr to be very small, which will lose the
then broadcasts a “request barrier edge” message, if it can
benefit of saving energy when sleeping, the barrier domino
get the responses from left and right neighbors in a waiting
phenomenon will be explicit during belt switch. RPM does
time, we get to the next step, or else A will backoff another
not need belt switch either, so its wake up scheme is the
time.
same as that of APM.
For the first free left neighbor B that is waiting for
command, if it heard A’s “request barrier edge” message,
and the first free right neighbor to be C. The B,C will send
A a “respond barrier edge” message. A finds out that B,C
are the first left and right neighbors to respond, it will reply
an “accept” message to B,C respectively. Then B,C become
A’s barrier neighbors. After this process, B,C become the
new barrier heads (see Fig. 6), and each of them will repeat
the above steps. But this time, B,C only need to receive one
“respond barrier edge” message before it hands off the role
of barrier head. After a node finds that there are k brothers
become the barrier heads, and it is still free, then it will be
scheduled to sleep.
6.2 Sleep Schedule Figure 7. Network Lifetime for Barrier Coverage
There is no need for an accumulation belt to use belt Deployment Models
switch. Even if multiple barrier spaces are considered, e.g.
( A23 , A24 , A12 ). First, for this system, we can use A59 to
7. Simulation Results the low quality sensors. This indicates APM is a very robust
and reliable structure. Even if the quality of the sensors are
To evaluate our analysis, we build a simulation platform not good and vulnerable to all kinds of failures, we can still
based on C++. We set the sensing radius R = 10m, maintain a good network service if we deploy sensors
communication radius Rc = 30m. Sensors are deployed in a according to APM.
belt region of 3000m, and for RPM and APM, the width of
the belt is 20m. As for IBM, the width is not fixed in order
8. Conclusions
to make sure different belt independent. The k is fixed to 3,
and t is a variable. The overlap factor θ = 0.1. The sensors’ Compared with existing models of barrier coverage -
mean lifetime μ = 400, the variance of lifetime σ = 20. We Random Point Model and Independent Belt Model, the
deploy different number of sensors, so that the node Accumulation Point Model is an economical, robust and
densities are different. Here the network energy residue rate reliable structure: its cost is the least, and the failure
ρr is a metric to measure how much energy remained when resistance is the best. The Accumulation Point Model is a
the barrier fails, long lived deployment model for sensor networks. Although
originated from barrier coverage, it is possible to use
Accumulation Point Model for area coverage. If we deploy
the accumulation points instead of single sensors, the fault
Fig. 7 indicates RPM needs more sensors to maintain a tolerance and network lifetime will definitely increase. This
certain network lifetime. When the number of sensors is will be more helpful than simply increase the node density.
lower than 1800, the network lifetime of RPM is zero. This
is because RPM needs a high node density to work. After the Acknowledgments
density reaches a threshold, RPM’s lifetime increase fast,
but it is still lower than IBM, while IBM’s network lifetime
This work is supported in part by the National Science
is lower than APM. In this experiment, we optimize the
Foundation under award number CT-ISG 0716261.
backup nodes and backup belts to sleep in order to save
energy and maximize the network lifetime. That is why the
network lifetime can be 3000 days for APM, nearly 9 years. References
[1] P. Balister, B. Bollobas, A. Sarkar, and S. Kumar.
Reliable density estimates for coverage and connectivity
in thin strips of finite length. In Proceedings of the 13th
annual ACM international conference on Mobile
computing and networking (MobiCom), pages 75–86,
NY, USA, 2007.
[2] Ai Chen, Santosh Kumar, and Ten H. Lai. Designing
localized algorithms for barrier coverage. In Proceedings
of the 13th annual ACM international conference on
Mobile computing and networking (MobiCom), pages
63–74, New York, NY, USA, 2007. ACM.
[3] Ai Chen, Ten H. Lai, and Dong Xuan. Measuring and
Figure 9. The Effect of Sensor Quality on Network Lifetime guaranteeing quality of barrier-coverage in wireless
sensor networks. In Proceedings of MobiHoc, pages 421–
The network energy residue rate in Fig. 8 is an indicator 430, NY, USA, 2008.
for the performance to price ratio. Low network residue rate [4] X. Cheng, X. Huang, and Xiang yang Li. Applications
indicates a High Performance to Price Ratio. APM’s residue of computational geometry in wireless networks, 2003.
ratio is the least, which means its waste is the least of all [5] Qunfeng Dong. Maximizing system lifetime in wireless
models. RPM has a higher residue rate than 50%. This sensor networks. In Proceedings of the 4th international
means when the network cannot meet the quality symposium on Information processing in sensor
requirement, only less than half of the whole energy networks (IPSN), page 3, Piscataway, NJ, USA, 2005.
consumed. The graph also indicates that when more sensors IEEE Press.
are deployed, the residue rate decreases as a tendency. [6] D.W.Gage. Command control for many-robot systems.
In the next experiment, we need to find out the network In AUVS-92, the Nineteenth Annual AUVS Technical
failure resistance ability of different deployment models. We Symposium, pages 28–34. Unmanned Systems
fix the number of sensors to be 3000, and change the sensor Magazine, 1992.
lifetime variance σ from 15 to 130 days, in order to indicate [7] J.A. Fuemmeler and V.V. Veeravalli. Smart sleeping
we use sensors with poor quality. All the other parameters policies for energy efficient tracking in sensor networks.
are the same as the first experiment. We find out that IBM’s In IEEE Transactions on Signal Processing, pages 2091–
lifetime deteriorates fast from 1800 to 0, even though the 2101, Washington, DC, USA, 2008. IEEE Computer
number of sensors remains the same (Fig. 9). As for RPM, Society.
although its lifetime is not long, only 390 at the beginning, [8] Rajagopal Iyengar, Koushik Kar, and Suman Banerjee.
it does not decrease so sharply as that of IBM. APM’ Lowcoordination topologies for redundancy in sensor
lifetime is the longest. It only decreases a little when using networks. In Proceedings of the 6th ACM international
(IJCNS) International Journal of Computer and Network Security, 31
Vol. 2, No. 10, 2010
Author’s Profile