Patterns in the RSSI Traces from an Indoor Urban Environment Nicholas M. Boers, Ioanis Nikolaidis, Pawel Gburzynski Department of Computing Science University of Alberta Edmonton, Alberta T6G 2E8 Canada Email: {nboers, nikolaidis, pawelg}@ualberta.ca Abstract—Urban environments are notorious for their high spectrum usage, particularly in their unlicensed radio bands. Wireless sensor network (WSN) nodes incorporate modern transceivers that can measure the background noise/interference and change channels. These combined capabilities suggest the need to better understand urban environments so that nodes can better avoid competing devices. In this paper, we explore the noise and interference patterns found on 256 frequencies in an indoor urban environment’s 900 MHz ISM and non-ISM bands. We begin the process by using off-the-shelf WSN hardware to sample the environment at 5 kHz from 16 locations simultaneously. From these samples, we identify five prevalent patterns and then hand-classify the 4096 traces of noise and interference. Finally, we extract a variety of statistics from the traces and use them in a Bayesian network classifier. I. I NTRODUCTION Dense and dynamic urban environments present both op- portunities and challenges for wireless sensor network (WSN) deployment. The availability of electrical outlets and backbone wired networks allows for less dependence on batteries and fewer hops, respectively. On the other hand, the wireless de- vices typically operate within one of the industrial, scientific, and medical (ISM) radio bands. These bands are heavily used, e.g., by cordless phones, wireless local area networks, building automation networks, and microwave ovens, so devices need to be resilient to interference. When a wireless node receives a transmission, the ratio between its signal strength and any interference plus back- ground noise (SINR) ultimately determines its fate [1]. In an environment without motion, received signal strengths tend to remain relatively stable over time. Background noise is often assumed to be additive white Gaussian noise (AWGN), and like received signal strengths, remains relatively stable. The final variable, interference, is a potentially large uncontrolled source for variation in the SINR, and yet, interference has generally received little attention in the literature. Strong interference can affect all of the nodes in an environment, and channels operating near their receive sensitivity are especially sensitive to even small changes in the SINR. During experiments with a medium-scale indoor WSN [2], we found that nodes often experienced high packet losses even at short distances and periods of no congestion. These difficulties led us to this exploration of their behaviour at that frequency and many others. Our work adds to the already healthy scepticism related to the proper modelling in network simulations that was started in earnest with the paper by Kotz et al. [3]. However, we target specifically the problem of characterization based on evident interference patterns. After sampling the transceiver’s received signal strength indicator (RSSI), which measures its radio-frequency power input, on 256 different frequencies, we aim to (a) highlight the importance of evaluating channels prior to performing real- world experiments and (b) show that while many channels may be suitable for low-powered communication, some clearly are not. Recognizing these poor channels is important because modern transceivers have the ability to change channels. When nodes detect that they are using a poor channel, they can make the change to a different frequency. We have no illusions about the generality of our results: we sampled one specific environment, on a particular date, over a particular period of time. We realize that the same environment may exhibit different characteristics if sampled again, and other environments may be completely different. That said, it is important for researchers to realize that the combination of noise and interference is rarely straightforward AWGN. Our work is relevant to the area of cognitive networking, and in particular, it attempts a high-and-level characterization of channels that are potentially occupied by (licensed or unlicensed) users. In cognitive radio, the characterization is performed with the intent of determining the (near-term) occupancy as expressed through the observed noise and inter- ference combination. Cognitive networking needs such a step before opportunistically using one or more of those channels. In fact, rather than producing a binary (occupied/unoccupied) classification of channels, we characterize channels into five different classes. Some of the classes, for example those ex- hibiting presence of spread spectrum or ultra-wideband inter- ferers, could, depending on the cognitive networking scheme, be used at the same time for narrowband transmissions by the cognitive network. However, in our study, we do not take a position as to how the channels will be used, but rather what is a good (and as exhaustive as possible) characterization of channels based on their noise and interference time series 2010 15th IEEE International Workshop on Computer Aided Modeling, Analysis and Design of Communication Links and Networks (CAMAD) 978-1-4244-7635-0/10/$26.00 ©2010 IEEE 61