Model-Based Event Detection in Wireless Sensor Networks

A number of testbeds (e.g., [1][2][3]) have shown the potential of wireless sensor networks (WSNs) to collect environmental data at previously unimaginable spatial and temporal densities. These developments present many data management challenges. First, our experience from the deployments has made clear the shortcomings of the static behavior of current sensor networks. For example, scientists would like to sample the environment at a high frequency to capture detailed information about "interesting" events, but doing so would create an inordinate amount of data. On the other hand, sampling at a lower frequency generates less data but misses important temporal transients. Second, the large amount of data that these networks generate complicates the querying and post-processing stages. Rather than manually traversing through the collected data, scientists would prefer to query for measurements related with certain events (e.g., significant rainfall).

To address these issues, we need WSNs that can reason about the phenomena they observe and change their behavior based on events they detect. Possible adaptation strategies include changes in the sampling rate as well as waking up other nodes in the network to increase spatial coverage of the detected event [4,5].

The readings of sensors are superpositions of several processes. They are often dominated by predictable foregrounds, which can be very much larger than the subtle trends and variations that we are trying to measure or the small events that we try to detect. In order to interpret the readings, it is important to separate these different signals into independent components. In environmental monitoring, most sensors witness daily variations of all quantities and seasonal trends. In addition, there are discrete natural events (storm, rainfall, strong winds) that have a separable effect on our data. We present an approach using techniques of statistical signal processing to decompose the sensor readings into various physically meaningful components. In our approach, we perform a step-by-step identification of various foregrounds. We identify the diurnal cycle present in both the box and soil temperature sensor data and we account for the effect of seasonal drift. We make use of all these priors (daily cycle, seasonal drift) to detect events by identifying when measurements diverge from those expected by the foregrounds.

Specifically, we explore variants of Principal Components Analysis (PCA) [6] that we use to extract features from the data collected by the network and discover the multiple underlying physical processes that generate the observed data. This produces a model of “normal behavior.” Observations that diverge from the model correspond well with events. We note that one can build the PCA model offline using historical data and that a small number of parameters summarize the phenomena that the motes sense. Such a compact representation of the model makes it possible to build a lightweight event detection mechanism that runs in real time on the network’s motes.

We evaluate the performance of the proposed mechanism using data from the Life Under Your Feet environmental sensing network [1]. We execute the event detection algorithm to detect rain events with the deployment area over ten months of the network’s lifetime. We compare the list of detected events with precipitation data recorded by a weather station at BWI airport.

This specific application reveals another aspect of the proposed approach: while the motes in our network have soil moisture sensors, these sensors cannot detect the onset of a rain event, because soil moisture rises only after the water seeps through the soil. Instead, we use a combination of air and soil temperature measurements to detect when rain starts to fall. Figure 1 shows that temperature varies immediately with the onset of an event, but that soil moisture lags by several hours. The model allows us to detect the rain event rapidly based on indirect evidence prior to the rain’s direct effect on soil moisture. This better describes system behavior, capturing much more information about the dynamics of soil moisture in response to rain.

While our solution is generally applicable to WSNs that collect large amounts of data using multiple sensing modalities, we present our design through a environmental monitoring application we developed and is currently deployed for over 18 months at an urban forest in Baltimore, MD. The purpose of the Life Under Your Feet network is soil monitoring in which each of the network’s ten motes periodically collects measurements, including soil temperature and soil humidity, as well as ambient temperature and light.

The key difference between this application and previous environmental monitoring networks (e.g., [2,3]) is that all raw measurements are reliably retrieved at the network’s base station, which subsequently inserts them to an SQL database. This stringent reliability require

…(Full text truncated)…

This content is AI-processed based on ArXiv data.