Stream Search brings search technology to streams of sensor data. Existing search technology is based primarily on text and hyperlinks, but sensor data is a challenge for search because it is primarily numeric. We are using data mining techniques to create the algorithmic equivalant of hyperlinks between data streams.

The first type of link that we are creating between data streams is based on temporal correlation: streams that are temporally correlated are somehow related. Part of our goal is to create definitions of temporal correlation between infinite streams of data that (i) are computationally efficient (ii) provide bounded translation invariance, and (iii) work for both digital and analog data streams.

Preliminary results indicate that temporal clustering can be used to identify patterns, which can be translated into correlations between data streams. We are testing current techniques on data collected from instrumenting people's homes, including digital sensors on doors, motion, sinks, microwaves, showers, toilets, cabinets, and couches. We are also using analog data from accelerometers worn by home residents. This technology will soon be applied to sensor streams generated by the MetroNet project and the VineLab testbed.

Robert Dickerson, Jiakang Lu, Jian Lu, and Kamin Whitehouse. Stream Feeds - An Abstraction for the World Wide Sensor Web. Conference on the Internet of Things (IOT) 2008. March 26-28. Zurich, Switzerland.