You are on page 1of 6

A Topology Inference Algorithm for Wireless

Sensor Networks
Theofanis Kontos1, George S. Alyfantis1, Yiannis Angelopoulos2, Stathes Hadjiefthymiades1
1

National and Kapodistrian University of Athens; 2Athens University of Economics and Business

Abstract The use of network tomography has been proposed


as a means to infer network topology in wireline networks. In
what regards wireless sensor networks, the methods suggested
for topology discovery make explicit use of neighborhood
information sent by internal network nodes. We propose an
algorithm for inferring the tree topology of a wireless sensor
network based on loss measurements collected at the sink of the
network. The proposed method does not rely on the active
cooperation of internal network nodes, but is capable of inferring
the sought topology, only through the received sensor readings
(passive method), without undertaking energy-intensive tasks
within the network.
Index Terms Ad hoc networks, sensor networks, network
tomography

I. INTRODUCTION

IRELESS sensor networks consist of nodes with


minimal resources in terms of computing power,
bandwidth, storage and energy. Task scheduling schemes and
network management operations can assist in rationalizing
resource utilization but they typically require topology
information. Topology discovery in WSNs requires the active
co-operation of network nodes rendering most schemes
energy-intensive.
In this paper, we address the problem of passive topology
inference of WSNs involved in tree-based data collection
tasks, assuming minimal cooperation from the internal
network nodes. The proposed algorithm is based solely on
measurements received at the sink, thus, providing a passive
topology inference method. Specifically, the algorithm is
based on the observed losses and their correlations that reveal
topology information. The algorithm caters for a convergecast
environment or reverse multicast tree setting [10], [11] where
the network nodes are involved in a common data collection
task with periodic information transmission [8]. The algorithm
gradually reveals the sought tree topology in increasing degree
of detail, on a per measurement basis.
In WSNs, such operations are usually of a distributed
nature, based on local information and local decisions. We
propose a method to discover the global topology, since this,
combined with knowledge of dynamic factors like resources,
link quality, user behaviors, and application data could be of
use for network management. Our algorithm helps WSNs
deployed on periodic data collection tasks to address this issue
taking into account their resource scarcity.

978-1-4673-2713-8/12/$31.00 2012 IEEE

The rest of the paper is organized as follows. Section II


summarizes relevant previous work. Section III outlines our
assumptions and describes the adopted model. We continue
with the presentation of the proposed WSN topology inference
algorithm, in Section IV. In Section V, we present results from
the algorithms assessment. Finally, in Section VI, we discuss
our conclusions and provide some suggestions for future work.
II. PREVIOUS WORK
The problem of topology discovery has been extensively
studied for the Internet [7], or smaller wireline networks [9].
The so called tomographic network topology inference
methods ([1], [3], [4], [7]) focus on the observation of
measurable quantities at the edges of the network. They can be
further separated into active or passive depending on whether
explicit cooperation from internal nodes is required. Considerably less work has been performed for the topology
discovery of wireless sensor networks.
WSN topology was first inferred using active methods,
where nodes are polled on information on their neighborhood.
Usually a subset of privileged nodes perform this task. The
TopDisc algorithm [6] is an application of this method, as well
as the methods in [2] and [5]. Besides the advantage of
topology-independent validity, the energy cost associated with
active methods is a drawback.
In [9] a tomographic approach is employed to infer the
topology of a modest-sized sensor network utilizing
expectation maximization (EM), which has been already used
for wireline networks and is a well-accepted approach.
More recent contributions rely on the observation of end-toend losses to perform the WSN topology inference task. These
methods tend to be more passive, with the inferring algorithm
running at the sink [12]-[15]. In [15], loss is utilised by the
inference algorithm and in [14] the correlation between data
received from different sensor nodes serves the same purpose.
[13] suggests the probabilistic transmission of neighborhood
information and use of EM at the sink to infer global topology.
The algorithm we propose is also tomographic and passive
as it is based on measurements received at the sink for the
purposes of other applications and internal nodes need not be
aware of the topology inference process. It provides a
technology-agnostic solution, where initial information, such
as number of nodes, connectivity or location information, is
not needed a-priori, as input to the algorithm.

000479

III. SYSTEM MODEL


Let us assume a WSN involved in a periodic data-collection
task, while implementing a convergecast topology [11] with
the reverse multicast tree [10] rooted at the sink of the WSN.
Playing a role in such a task, each node aggregates the
readings received from all its children and forwards them to its
parent. This is performed at every level, until the sink is
reached. Let T{V, E} denote the tree that represents the WSN,
where V= {1,,S} is the set of sensor nodes (vertices), while
E is the set of links (edges) connecting the nodes. The link
connecting node j with its parent i is represented through the
pair (i,j). This parent-child relationship is also denoted with
pj=i. It is assumed that for the time that it takes for the
inference to be completed the topology remains unchanged.
Each node has the capability of collecting and routing data
back to the sink via a wireless multi-hop infrastructureless
routing protocol, appending its own id to the datagram
transmitted. The result received at the sink is a dataset with a
trail of node identities and some aggregated data.
The sink receives results (datasets with a trail of node
identities) at regular time intervals named epochs. Due to
possible link losses, some node identities may not appear in a
dataset. Note that, due to the aggregation process, if a link
fails the ids of the respective node and all its descendants are
not received at the sink. This can happen if a node fails to
transmit and also the waiting period at its parent node expires.
This waiting period is predefined and corresponds to an epoch.
Message losses take place at every link with a certain
probability. The vector of link loss probabilities is denoted
with Q and it holds: Q = (q1,q2,,qS). The data collection at
every epoch can be then perceived as an experiment, the
outcome of which is a binary vector of the form mk = (mk,1, ,
mk,S), where k denotes the kth measurement, S denotes the
number of nodes of the network1, and
1, if data received from node
mk ,a =
0, otherwise

(1 )

This vector can be readily constructed at the sink, after the


reception of a dataset. The sensor readings themselves are not
of interest to the algorithm.
IV. TOPOLOGY INFERENCE ALGORITHM
In this section, we describe a method for inferring the
topology of a WSN starting with a totally abstract view of the
WSN topology. The proposed algorithm is based on the notion
of node clusters (NC). A NC is a subset of the sensor nodes
comprising the sought topology. The relationships among its
member nodes are originally unknown. NCs are
interconnected, so as to form a tree, representing an abstract
view of the real tree topology. NCs are gradually broken up
and placed at the right position in the tree, until all NCs have
finally become singleton sets. Let T{V, E} be the tree
1

Note that, in fact there is no need for initial knowledge of the node count,
S. The proposed algorithm may readjust (expand) the size of vector mk every
time a node appears for the first time in a measurement. Section IV describes
how the algorithm handles this issue. Here, S is used for notational simplicity.

978-1-4673-2713-8/12/$31.00 2012 IEEE

topology that we want to infer. At any time, there is at least


one NC, but no more than S NCs. Therefore, if I is the set of
indices representing different NCs, then it holds that I {1,
, S}. Moreover, NCs are disjoint sets, i.e., NCi NCj = ,
i, j I and their union forms the set of nodes of the tree, i.e.

U NC

=V

(2 )

iI

As already stated, the proposed algorithm is based upon the


presence of the ids of descendant nodes observed at the
received measurements. Based on equation (1), let Li = {j
NCi | mj = 0}, i I, be the set of nodes of NCi that did not
appear in the last measurement (lossy nodes). We then define
a measure of loss for NCs as follows:

li =

Li
NCi

(3 )
, i I.

li assumes values in the interval [0,1]. If li = 1, we


characterize NCi as fully lossy, while if 0 < li < 1 as partially
lossy. Lastly, if li = 0, we characterize NCi as lossless.
The main idea of the algorithm is that if a node i is lossy in
one epoch and a node j is not, then i cannot be an ancestor of j.
An example is shown and explained in Fig. 1(a).
The proposed algorithm breaks down to three distinct steps
in order to process a received measurement:
Node cluster splitting - the main idea
Merging and reordering node clusters using
introduced constraints
Invalid tree repair
In the following we elaborate these steps.
Node Cluster Splitting
Given a new measurement vector, some of an NCs nodes
appear in it, while others may not (i.e., 0 < li < 1). We split
such a partially lossy NC into two subsets (fully lossy, and
lossless) and make the first child of another NC (or pushing it
deeper into the hierarchy), while keeping the second where the
initial NC was. The logic behind it is that if a node fails, none
of its descendants should appear in the measurement. Hence, a
reasonable and simple solution would be to keep the lossless
sub-NC in its initial position, and make the fully lossy sub-NC
its child. The process of NC splitting is depicted through an
example in Fig. 1 and formulated in Listing 1. This splitting
can only be applied to partially lossy NCs.

Fig. 1. Process of NC splitting. Received measurement: 11100 (i.e., nodes 4


and 5 not present). (a) tree before splitting of the partially lossy NC. Lossy
nodes 4 & 5 cannot be ancestors of non-lossy 1 & 3. (b) tree after splitting the
partially lossy NC.

000480

Listing 1. PROCESS OF NC SPLITTING


Foreach i I, where 0 < li < 1 {
select j from the set {1, , S} I
set I := I {j}
set NCj := Li
set NCi := NCi Li
set pj:= NCi
}

1.
2.
3.
4.
5.

However, nodes may have been falsely separated into


different branches of the topology, which means that it may be
required to merge NCs, or reorder them. We introduce an
auxiliary data structure which assists in selecting the suitable
merges or reorderings among the numerous possible ones.
Introducing Constraints
Let P be an SS matrix, as the aforementioned auxiliary
structure. The elements of the P matrix represent constraints
on the relationships between every pair of nodes. Specifically,
(4)
1, a not a descendent of b

Pa ,b =
0,

otherwise

Before the receipt of any measurement, we have no


indication of such constraints (theoretically, any node can be a
descendant of any other node), i.e., Pa,b = 0, a, b V.
However, upon receipt of a measurement, say mk, we may
elicit such constraints. Specifically, we may set Pa,b = 1, a {j
V | mk,j = 1}, b {j V | mk,j = 0}. In other words, if a node
a is present in a measurement, while b is not, because of the
tree topology assumed, it is evident that a cannot be a
descendent of b.
Merging and Reordering Node Clusters
Based on matrix P, we may now consider the process of
merging/reordering. Let NCi be the NC that is currently
studied. We take each of its siblings, i.e., NCs from the set
{NCj, j I | pj=pi}, and check their relation to NCi with
regards to the P matrix2. The possible relations with regards to
the P matrix are the following:
1. Pa,b = 1, a NCi, b NCj
2. Pa,b = 0, a NCi, b NCj
3. Pa,b = 1, a NCj, b NCi
4. Pa,b = 0, a NCj, b NCi
5. (a,b) NCiNCj | Pa,b = Pb,a = 1 (i.e., there are pairs of
nodes from NCi and NCj, which are mutually incompatible)
6. None of the above. No clear relation between all nodes
from NCi and all nodes from NCj
According to their relation, one of the following actions can be
performed:
a. Merge NCi with NCj,
b. Make NCi child of NCj,
c. Make NCj child of NCi, or
d. Do nothing

relation between and NCi and NCj, we may finally take one of
the following actions:
Action b., if both 2. and 3. hold true.
Action c., if both 1. and 4. hold true.
Action d., if both 1. and 3 hold true, or in case 5. is true.
Action b. (or c., or a.), if both 2. and 4. hold true.
Action a., in case of relation 6.
The process of merging and reordering is formulated in
Listing 2. Examples of the aforementioned actions are
depicted in Fig. 2.

Listing 2. PROCESS OF MERGING AND REORDERING.


Merge NCi and NCj:
1. NCi := NCi NCj
2.
NCj :=
3.
I := I {j}

Make NCj child of NCi:


1.
pi:= NCi

(a) Merging two NCs


(b) Reordering two NCs
To summarize, when we examine two NCs, if we know that
all the nodes of the first could be descendants of the second,
while no node from the second could be descendent of the
first, we make the first NC child of the second. Similar logic
serves the complementary case. In the case that no node from
the first NC can be a descendant of the second NC and vice
versa (or even if there is a pair of mutually incompatible
nodes), no action is taken, so as to keep the topology valid.
When all nodes from the first NC can be descendants of the
second NC and vice versa, i.e., there is no constraint between
the nodes of the examined NCs, we may either merge the two
NCs into a single NC, or make any of them child of the other.
Last, if no clear relation between the examined NCs has been
identified, we merge these NCs into one.
Invalid Tree Repair
This important step serves toward the identification of
invalid topologies and their repair. A topology is not valid, if
there is at least one partially lossy, or lossless NC, say NCi,
(i.e., li < 1), which has a fully lossy ancestor NCj (i.e., lj < 1).
In such a case, we need to take corrective actions, to overcome
the determined flaw. Let Pi ( I) denote the set of ancestors of
NCi and Fi denote the set of fully lossy ancestors of NCi, i.e. Fi
= {j Pi | lj = 1}.
If li < 1 (i.e., NCi is partially lossy or lossless), and
Fi (i.e., there is at least one ancestor of NCi that is fully
lossy), the topology is not valid. In this case, we may repair
the topology by shifting NCi to another place in the tree, such
that the resulting topology becomes valid. A reasonable choice
is to move NCi upwards into the tree and make it a sibling of
its higher-level fully lossy ancestor. This action is depicted in
Fig. 3. Observe that, in the plotted example, the NC containing
node 3 is fully lossy, while its child NC containing nodes 4
and 5, is partially lossy, which is an invalid topology. In order

According to the case to which we have classified the


2

The process of sibling examination is performed exhaustively, i.e., we


examine a NC with all its siblings. Then we examine the following NC with
the rest of the siblings, and so on.

978-1-4673-2713-8/12/$31.00 2012 IEEE

000481

ITR(NC(i), M) // Invalid Tree Repairing. Move NC(i), if necessary to


make the tree valid
EndFor
J := BTS(T) // Take the Breadth-first search sequence of NCs
For i in J
T := getSiblings(NC(i)) // Take the siblings of NC(i)
For k in T
// Check tvalidity of 6 different cases for each pair of NCs
isCase1 := isCase(1, NC(i), NC(k))
isCase2 := isCase(2, NC(i), NC(k))
isCase3 := isCase(3, NC(i), NC(k))
isCase4 := isCase(4, NC(i), NC(k))
isCase5 := isCase(5, NC(i), NC(k))

Fig. 2. Merging and reordering of NCs. (a) initial topology (b) the rightmost
NC becomes child of its leftmost sibling NC (c) the leftmost NC becomes
child of its rightmost sibling NC (d) the two sibling NCs are merged into a
single NC.

to correct such inconsistency, we make the child NC a sibling


of its (fully lossy) parent. The repairing process is performed
for all NCs of the tree, which are examined in a breadth-first
way. Listing 3 illustrates the discussed process. In this listing
dj denotes the depth of NCj.

Listing 3. REPAIRING AN INVALID TREE (NC EXAMINATION)


Invalid tree repair (examination of NCi):
1. If li < 1 and Fi , go to 2. Else go
to 3.
2.
3.

p i := arg min d
j F i

Finish

Operation of the WSN Topology Inference Algorithm


The proposed topology inference algorithm, at each
iteration (measurement), combines the three steps described
above. However, for presentation reasons, the order by which
they have been discussed is not their actual execution order. In
the proposed algorithm, upon receipt of a measurement, the
algorithm first checks if the tree is valid, and if not, takes
corrective actions. In the sequel, it checks sibling NCs in order
to make all possible merges/reorderings, and, lastly, performs
NC splitting, where possible. In Listing 4, the full algorithm is
described in pseudocode.

isCase6 := isCase(6, NC(i), NC(k))


If (isCase2 AND isCase3)
Move(NC(i),NC(k)) // Make NC(i) child of NC(k)
Else If (isCase1 AND isCase4)
Move(NC(k),NC(i)) // Make NC(k) child of NC(i)
Else If (isCase2 AND isCase4)
Move(NC(i),NC(k)) // Make NC(i) child of NC(k)
Else If (isCase6)
Merge(NC(i),NC(k)) // Merge NC(i) with NC(k)
EndIf
EndFor
EndFor
// Take the Breadth-first search sequence of NCs
For i in J
If (isPartiallyLossy(NC(i), M)) // If NC(i) is partially lossy
Split(NC(i), M) // Split the NC
EndIf
EndFor
EndFor

It has to be noted that in the description of the algorithm, it


is implied that the number of nodes is known. Note, however,
that this has been assumed for the sake of presentation, as it is
not necessary to know a priori such information. The
described algorithm can perform dynamic discovery.
Specifically, upon receipt of the first message from a node,
this node can be included in a child NC of the sink. It may
later be moved within the tree depending on the subsequently
received measurements.

Listing 4. PSEUDOCODE OF THE MEASUREMENT PROCESSING


OF THE PROPOSED TOPOLOGY INFERENCE ALGORITHM
// Create a NC and populate it with all nodes of the WSN
For a=1:S
NC(1).add(a)
EndFor
p(1) = NC(0) // NC(0) represents the sink. The only NC, NC(1), is set as child
of the sink
For each M // For each measurement
UpdateConstraints(M) // update the constraints matrix, based on the last
measurement
For i in J

978-1-4673-2713-8/12/$31.00 2012 IEEE

Fig. 3. Repair of an invalid topology. (a) Invalid topology (b) Corrected


topology

We assume the following stopping conditions for the


algorithm:
The reconstructed topology is composed of singleton
NCs, and

000482

The P matrix indicates that each node can be a


descendant of a set of nodes, which are indicated as its
ancestors in the inferred topology. Let

Ai =

U NC

(5 )

j Pi

be the set of sensor nodes that are ancestors of NCi in the


reconstructed tree, and

Bi = {b V Pab = 0, NCi }

(6 )

the set of sensor nodes that can be ancestors of NCi, according


to the P matrix. Then the test that can be adopted to see if the
reconstructed topology is the final one is to check if the
following holds true:

Ai = Bi , NCi = 1, i I

( 7)

network Ti) is in their parent node cluster and PNF-i is the


respective percentage.
Fig. 4 and Fig. 5 show our findings. In Fig. 4 we show the
number of epochs required for full topology discovery against
the link failure probability. Each curve in the depicted family
of curves corresponds to a value of the number of network
nodes. The number of epochs needed for the topology to be
inferred correctly (I#) is minimum for a span of modest values
of q (between 0.02 and 0.25). For very low and very high loss
probabilities, I# assumes greater values; in the absence of loss,
there is no input for the algorithm and in case of high losses it
is harder for nodes to differentiate themselves from their
ancestors. In Fig. 5(a), we observe that Ci (the percentage of
the topology that is inferred after the i-th measurement)
reaches 0.9 much earlier than the full topology is inferred,
which means that fairly accurate reconstructed topologies are
produced quite quickly. The fluctuations in the Ci metric,
depicted in Fig. 5(a), are caused by the corrections performed
by the algorithm (see Section IV).

V. SIMULATIONS
We evaluated our algorithm with a custom made simulator.
We used the popular JGraph library for J2SE to handle the
evolving network graph throughout the topology discovery
process.
The set of sensor nodes V and the set of link loss
probabilities Q which is used to create test data are the input
parameters for the algorithm. A topology T is generated
randomly according to V and additionally for every iteration i
a measurement mi is generated according to Q. Note that our
model does not assume that link loss probability is the same
for all links. However, in the simulations we follow the
simplifying assumption that Q = (q1,q2,,qS) = (q,q,,q) to
serve our proof of concept. In our experiments we try to infer
topology T based solely on the measurements generated and
we test the convergence of the algorithm by comparing it to
the true topology of the network, T that was initially
randomly generated.
Let Ci denote the percentage of the topology that is inferred
in epoch i. Ci is calculated according to (8).

Ci =

PNF i
PNC i

(8 )

where

NC i#
S

(9 )

NFi #
=
S

(10)

PNC i =
and

PNF i

where NC#i is the number of node clusters at the i-th epoch,


PNC-i is the percentage of the tree nodes inferred after the i-th
epoch, NF#i is the number of nodes in the inferred topology
whose actual father (according to the true topology of the

978-1-4673-2713-8/12/$31.00 2012 IEEE

Fig. 4. Measurements needed for different values of q and S.

One may perform some back-of-the-envelope calculations


for the convergence speed of the algorithm, providing some
initial insight on its performance boundaries. As a lower
bound, consider the following scenario. If we start with a
single NC of S otherwise ungrouped nodes, we may, in a
somewhat simplifying manner, assume that at every epoch the
lossy nodes are detached to form a new NC, hence doubling
the number of NCs. This will allow the NC tree to degenerate
to a tree of singletons no sooner than after log2S epochs. On
the other hand, a more pessimistic scenario is that at every
epoch only a few (only one in the worst case) nodes in every
NC are lossy. Therefore, the new NC formed of the lossy
nodes will contain only one node, thus, reducing the
population of the original NC by one. A simplistic conclusion
is that (S-1)/q epochs will be needed till all NCs become
singletons. A complementary analysis serves the case where
all but one node are lossy in every NC. Hence, according to a
simplifying estimation, the number of epochs required for
inference falls between a lower and an upper bound, which are
O(log2S) and O(S), respectively. In Fig. 5(b) we use the same
data as above, to display the number of epochs required for
topology inference against the number of nodes. The behavior
of the I#(S) function is monotonically increasing, albeit slower
than a linear one g(S)=(S-1)/q.

000483

VI. CONCLUSIONS AND FUTURE WORK


In this work, we have presented a passive, technologyagnostic algorithm for inferring the topology of a WSN based
on end-to-end loss measurements performed at the sink of an
inverse multicast tree. The novelty of this method lies on that
it does not rely on the active contribution of the internal WSN
nodes.

sink not only obtains the global topology, but also the network
status such as residual energy or link SNR. The result can be
conceived as a graph with an energy or SNR value tag
associated to each vertex. Such awareness of the global
network status would leverage network management
operations.
REFERENCES
[1]

[2]

[3]

[4]

(a)
[5]

[6]

[7]

[8]

(b)
Fig. 5. (a) Percentage of topology discovery according to the number of
measurements for constant values of q and S. (b) Number of epochs required
for topology inference against number of nodes S for different link loss

[9]

[10]

probabilities q

Another important characteristic of the proposed algorithm


is that the sought topology is inferred on-the-fly, with
increasing level of detail, as link loss measurements are being
fed to the algorithm. As a result, it is possible to trade between
speed and precision if desired. The assessment of the
algorithm through simulations has also shown that the
proposed scheme is quite efficient, regarding its speed of
convergence.
An interesting direction for future work would be to relax
certain assumptions, and consider that aggregated messages
may become long as they approach the sink. The enhancement
of our scheme to cover settings with time-varying topologies
and inhomogeneous loss across the network is also interesting.
The speed of discovery of subtrees rooted at a slow node (link)
is decided by the speed of the slow root node. In an
inhomogeneous setting a slow link affects the speed of
discovery of the whole subtree rooted at it. One intuitively
expects the full topology to be inferred at the speed of its
slowest subtree. However, the tradeoff between speed and
precision can offer an answer to this challenge, depending on
the application.
Another challenge would be the integration of the algorithm
within a global resource management framework, which
would take advantage of global topology awareness in order to
achieve energy efficient network operation: by including extra
information to the datasets that travel uplink to the sink, the

978-1-4673-2713-8/12/$31.00 2012 IEEE

[11]

[12]

[13]

[14]

[15]

[16]

R. Caceres, N. Duffield, J. Horowitz, and D. Towsley. Multicast Based


Inference of Network Internal Loss Characteristics. IEEE Trans. on
Information Theory, 45(24622480), 1999.
R. Chandra, C. Fetzer and K. Hogstedt. A Mesh-based Robust Topology
Discovery Algorithm for Hybrid Wireless Networks. Proceedings of
AD-HOC Networks and Wireless, Sept. 2002.
M. Coates and R. Nowak. Network Loss Inference Using Unicast Endto-End Measurement. In ITC Seminar on IP Traffic, Measurement and
Modelling, September 2000.
M. Coates and R. Nowak. Network Delay Distribution Inference from
End-to-end Unicast Measurement. In Proceedings 2001 IEEE Int. Conf.
Acoust., Speech, and Signal Processing, May 2001.
B. Deb, S. Bhatnagar and B. Nath. STREAM: Sensor Topology
Retrieval at Multiple Resolutions. Telecommunication Systems, Volume
26, Numbers 2-4 / June, 2004
B. Deb, S. Bhatnagar, Badri Nath. (Topdisc) A Topology Discovery
Algorithm for Sensor Networks with Applications to Network
Management, In Proceedings of IEEE CAS workshop, 2002.
N. Duffield, J. Horowitz, F. L. Presti, and D. Towsley. Multicast
Topology Inference from Measured End-to-End Loss. IEEE Trans. on
Information Theory, 48:2645, 2002.
G. Hartl, B. Li. Loss Inference in Wireless Sensor Networks Based on
Data Aggregation, In Information Processing In Sensor Networks,
Proceedings of the third international symposium on Information
processing in sensor networks, Berkeley, California, USA 2004
D. Marinakis, G. Dudek. Topology Inference for a Vision-Based Sensor
Network, In Proceedings of the 2nd Canadian conference on Computer
and Robot Vision, 2005
Yongyi Mao, Frank R. Kschischang, Baochun Li, and Subbarayan
Pasupathy, A Factor Graph Approach to Link Loss Monitoring in
Wireless Sensor Networks IEEE Journal on Selected Areas in
Communications, vol. 23, issue 4, April 2005, pp. 820-829
S. Ratnasamy and S. McCanne, Inference of Multicast Routing Tree
Topologies and Bottleneck Bandwidths using End-to-end
Measurements, in Proc. IEEE Infocom99, New York, March 1999.
Ji Wenli, Gao Teng, Jiang Meiyin and Li Dan, "Researching Topology
Inference Based on End-To-End Date in Wireless Sensor Networks",
Proc. IEEE 4th International Conference on Intelligent Computation
Technology and Automation - Volume 02 (ICICTA 11)
Xing Kai, Cheng Xiuzhen, Chen Dechang and Du David Hung-Chang,
"Topology Inference in Wireless Mesh Networks" Proc. 4th
International Conference on Wireless Algorithms, Systems, and
Applications (WASA 09)
Zhao Tao, Cai WanDong, Li YongJun, "MPIDA: A Sensor Network
Topology Inference Algorithm" Proc. IEEE 2007 International
Conference on Computational Intelligence and Security (CIS 07)
Tao Zhao, Wangdong Cai and Yongjun Li, "Using end-to-end data to
infer sensor network topology" Proc. 2007 IEEE Internat. Symp. Signal
Processing and Information Technology (ISSPIT 07)
Zhi-Yong Zhang and Guang-Min Hu, "Data aggregation-based topology
inference for wireless sensor networks" in Proc. International
Conference on Apperceiving Computing and Intelligence Analysis, 2009
(ICACIA 09)

000484

You might also like