![](bg2.png)
example, if a node detects a periodic spike of inter-
ference sufficient to classify it as a channel with peri-
odic interferer, it might agree with the class identified
on the same channel by another node, but there is no
guarantee that the two nodes sense the same periodic
interferer.
The model assumed throughout this paper is that
each node independently decides on what is the na-
ture of the interference via a classification technique
(placing it in one of five classes) as outlined in (Boers
et al., 2010). To facilitate a comparison of the back-
ground interference as seen by different nodes, we
develop a technique to correct the lack of synchro-
nization across the samples collected by the differ-
ent nodes. The lack of synchronization is caused by
the absence of a global clock and the individual node
clock drift. The purpose of the paper is to study, from
collected empirical evidence whether, if, when con-
sensus is reached, it is indeed valid, i.e., it concerns
the same interferenceseen by all the nodesat the same
points in time. To this end, we examine whether, if
consensus exists, the levels of interference are com-
patible across the nodes, i.e., the small time scale
behavior is the same. For example two nodes with
a valid consensus characterizing the channel as hav-
ing periodic spikes may still perceive different noise
floors and variance of noise between spikes, making
the potential Signal to Noise Ratio (if a transmission
were to be attempted) drastically different from the
perspective of the two nodes. In short, we are study-
ing whether a simple class-based consensus can be re-
lied upon to represent the common reality across the
nodes of the same WSN.
The rest of the paper is organized as follows. In
Section 2, we review the related literature. In Sec-
tion 3, we investigate and present the data we use in
our study. Section 4 explains the methodologywe fol-
low to analyze the data. Our results are presented in
Section 5. Section 6 provides concluding remarks.
2 RELATED WORK
Researchers studying the impact of external interfer-
ence in urban environments concentrate on identify-
ing and classifying patterns of noise and interference,
as well as applications of related classification tech-
niques to cognitive networking.
Lee, Cerpa and Lewis (Lee et al., 2007) measure
noise traces in many different environments in order
to propose algorithms to simulate noise and interfer-
ence. From these traces they observed three main
patterns of interference, (i) rapid spikes, (ii) periodic
spikes and (iii) noise patterns changing over time.
Boers, Nikolaidis and Gburzynski (Boers et al.,
2010) measured noise and interference in a four-by-
four node WSN, at high sample rates. They ex-
tracted five dominant noise and interference patterns:
(i) quiet, (ii) quiet with spikes, (iii) quiet with rapid
spikes, (iv) high and level and the (v) shifting mean
pattern. Consequently, they classified them using a
Bayesian network classifier. Later, this work was ex-
tended by classifying two of the aforementioned pat-
terns locally at each node using single-node decision
tree classifiers (Boers et al., 2012b).
In cognitive-networking, known identified pat-
terns can be exploited to coordinate cooperative sens-
ing across the nodes of a WSN. The determined
noise and interference patterns for each WSN can be
utilized to build a distributed classifier. In such a
scheme, the WSN nodes cooperate with each other
to reach a consensus on a specific pattern, after a
number of iterations, by exchanging and combining
their sensing information. This aims to eliminate
the impact of deficient individual pattern classifica-
tions (Akyildiz et al., 2011). The notion of coopera-
tive sensing extends also to multi-hop cases whereby
the sensing results of nodes are forwarded over mul-
tiple hops in order to improve the classification accu-
racy.
Rather than develop a scheme that attempts to
combine sensed measurements from the nodes to
reach a classification result, we examine whether indi-
vidual per-node classification and a simple network-
wide consensus is sufficiently accurate, at least in
WSN networks deployed in a small space such as the
network that is the object of this study. Per–node clas-
sification is justified because we wish to generate con-
sensus (and possibly revise it over time) without un-
due burden on the sensor nodes in terms of transmis-
sions. The alternative would have been the collection
of background noise signal strength data from all the
nodes to a host/sink that performs elaborate computa-
tion to decide on the state of the channel. Clearly,
the collection of all background noise and interfer-
ence samples to a sink is unattractive as it represents
a high energy cost to transmit them. Instead, we con-
sider an architecture whereby each WSN node classi-
fies, in isolation, the state of the channel and then a
consensus is derived using a message from each node
that indicates just the determined class, hence reduc-
ing the volume of data that need to be exchanged.
3 THE DATA
In this study, we use the RSSI traces collected by
Boers et al. (Boers et al., 2010), across 256 chan-
SENSORNETS2014-InternationalConferenceonSensorNetworks
250