Data Mining In Sensor Networks

In pace with rapid development and increasingly maturation in communication, embedded computing and sensing technology, sensor capable of sensing, communication and computation appears in the world. Networks of such devices typically consist of tens or hundreds of small, power constrained nodes deployed in remote locations which they are expected to monitor for months or years at a time. How to process the continuous large data streams in sensor networks efficiently and how to find interesting knowledge in these streams become new challenge. Ubiquitous data mining(UDM) is concerned with this problem.This paper mainly studys the data mining technologies in sensor networks, including classification, association rule, clustering. This paper presents an algorithm designed to efficient construct a decision tree over distributed data streams in sensor networks. A numerical interval pruning approach is used for efficiently processing numerical attributes, and a probabilistic bound on the accuracy is guaranteed. Rainforest algorithm framework is also used in this algorithm which can significantly reduce the size of the dataset to be processed. In this paper, an algorithm of finding frequent items in distributed data streams of sensor networks is developed. At first, the algorithm finds local frequent itemset in each sensor, and transmits them to their parents to merge. Repeat this, until they reach a central node. At last, the central node combines them into a global frequent itemset, and produces association rules based on them. This paper describes a techniquefor clustering homogeneously distributed data in sensor networks. The proposed technique is based on the principles of the k-means algorithm. At first, the central node generates k centroids and broadcast in the network. Thereafter, each sensor node assigns each point in their local dataset to the nearest centroid, and transmits their local k clusters information to their parents node to merge. This process is repeated until the central node receives all the information of its child nodes. For each cluster, the central node recomputes the centroid as the average of data points assigned to it. If it doesn't meet with the stop condition, the algorithm will iterates the process from the start. In the end, based on the above algorithms, the paper realizes a distributed data mining prototype system in sensor networks. The theory and the technology have the general signification and it is easy to extend to some application fields.