Equidistant Piecewise function Approximation for neurocomputing based environmental monitoring in wireless sensor networks

Environmental monitoring performed by an Artificial Neural Network (ANN) in wireless sensor networks (WSN) has turned out to be a suitable application [1]. Its main advantage is high accuracy prediction of environmental parameters, such as temperature or humidity [2]. Although predictors reduce in general the transceiver energy, their corresponding algorithm requires high calculation effort that may nullify this benefit. In order to decrease crucial mathematical terms of the original ANN algorithm, this work focuses on simplification by means of Equidistant Piecewise function Approximation (EPA). Thus, we split up the sigmoid function, which is used as activation function inside the ANN network, into several equidistant segments. The function slope inside each segment is replaced by a linear equation approximation. This minimizes the overall energy consumption as the calculation effort is reduced distinctly. For validation, our proposal has been implemented on a TelosB [3] sensor node (SN) where detailed evaluation and analysis of the EPA based ANN predictor is performed.