Sensor networks comprise small electro-mechanical devices that communicate over a wireless network. These devices collect environmental data and send them to a remote base station. The main goal of a data collection scheme for sensor networks is to keep the networks database updated while saving the limited nodes energy as much as possible. To achieve this goal without continuous reporting, data suppression is a key strategy. The basic idea behind data suppression schemes is to send data to the base station only when the nodes readings are different from what both nodes and base station expect. One alternative of data suppression is to cluster the nodes, aggregate their data and send only a summary to the base station. We propose to group the nodes into spatially homogeneous clusters, which consider both the geographical distance and the similarity of measurements between the neighboring nodes. Through simulated experiments, we have concluded that spatially homogeneous clusters produce data summaries with a higher statistical quality if compared with the usual ordinary clustering methods. Since distributed clustering algorithms play an important role in energy-efficient data collection proposals for sensor networks, we present Distributed Data-aware Representative Clustering (DARC) algorithm and Data-Aware Distributed Clustering Algorithm (DA-DCA). DARC and DCA build clusters around clusters representatives, which are able to produce more homogeneous clusters than the usual clustering proposals. Then, they produce data summaries that estimate the nodes data with a smaller error if compared with the usual data-aware clustering proposals. Another important characteristic of data suppression schemes is their sensitiveness to aberrant readings, since these outlying observations mean a change in the expected behavior for the readings sequence. Transmitting these erroneous readings is a waste of energy. In this thesis, we present a temporal suppression scheme that is robust to aberrant readings. We propose to use a technique to detect outliers from a time series. Since outliers can suggest a distribution change-point or an aberrant reading, our proposal classifies the detected outliers as aberrant readings or change-points using a post-monitoring window. This idea is the basis for a temporal suppression scheme named TS-SOUND (Temporal Suppression by Statistical OUtlier Notice and Detection). TS-SOUND detects outliers in the sequence of sensor readings and sends data to the base station only when a change-point is detected. Therefore, TS-SOUND filters aberrant readings and, even when this filter fails, TS-SOUND does not send the deviated reading to the base station. Experiments with real and simulated data have shown that TS-SOUND scheme is more robust to aberrant readings than other temporal suppression schemes proposed in the literature (value-based temporal suppression, PAQ and exponential regression). Furthermore, TS-SOUND has got suppression rates comparable or greater than the rates of the cited schemes, in addition to keeping the prediction errors at acceptable levels.
