Data clustering in sensor networks using ART

Abstract

Wireless sensor networks may be deployed in an environment the intrinsic pattern of data is unknown and may not be amenable to statistical aggregation techniques. Moreover, the dynamic nature of the environment may generate large amount data which may be highly similar with few data values that may be noisy or may indicate interesting observation. The present work proposes resonance based clustering of data. The technique works in two phases, offline and online phase. Fuzzy ART is employed as initial clustering process which is performed in the pre deployment offline phase to detect natural sets in data. Fuzzy ART does not require knowledge of the nature of data but in the scheme, the a priori knowledge of the type, range and other parameters are utilized to generate synthetic data to be used for clustering .The small number of data clusters eliminates the need to transfer large amounts of data. To cater to the dynamic nature of the environment, Fuzzy ARTMAP neural network (FAMNN) is employed at the cluster heads in the online phase to determine the group to which sensed data arriving after the formation of groups by ART. FAMNN is able to segregate outliers or form new groups whenever required. This obviates the need of ab initio regroupings or does not force data in one of the existing groups. To test the efficacy of the techniques, Fuzzy ART and FAMNN based clustering was performed and tested on synthetic sensor data with different parameters. Simulation results show that Fuzzy ART and FAMNN are able to identify the natural clusters and map new data to existing clusters or form new clusters to drastically reduce the amount of data required to be sent to the sink.