Implementing artificial neural-networks in wireless sensor networks

The development of wireless sensor networks is accompanied by several algorithms for data processing which are modified regression techniques from the field of multidimensional data series analysis in other scientific fields, with examples like nearest neighbor search, principal component analysis and multidimensional scaling (Guestrin, C. et al., Proc. IPSN´04, 2004). We argue that some algorithms, well developed within the neural-networks tradition for over 40 years, are well suited to fit into the requirements imposed by sensor networks: simple parallel distributed computation; distributed storage; data robustness; auto-classification of sensor readings. As a result of the dimensionality reduction obtained easily from the outputs of neural-network clustering algorithms, lower communication costs, and thus bigger energy savings, can be obtained. We present two possible applications of the ART and FuzzyART algorithms, which are unsupervised learning methods for clustering or categorization of the sensory inputs, applied on data obtained from a set of 5 Smart-It units (sensor nodes or motes) equipped with 6 sensors each. Results from simulations of purposefully faulty sensors show that these architectures are data robust to errors