A multisensor prediction-based heuristic for the internet of things

IoT devices usually produce large volume of data in low semantic levels. Also, they constantly communicate to the sink node. Since IoT devices are resource-constrained in terms of battery, memory and processing power, the usage of advanced machine learning techniques to support real-time decision making is unfeasible. So a challenge is to provide machine learning techniques tailored to the resource-constrained environment of IoT either by adapting existing techniques or developing new ones. In this paper, we propose an heuristic for adapting data prediction and data fusion techniques to preprocess data to avoid unnecessary communication between sensor devices and sink node. Also, in this work we compare (i) linear estimation; (ii) Weightless Neural Networks; and (iii) Moving Average Convergence Divergence in the aforementioned context. The main contribution is the proposal of an heuristics to mitigate unnecessary communications while avoiding data accuracy loss by combining data fusion and data prediction techniques. As implementation of the heuristics we present in this work, we introduce two algorithms: Dione and Delfos. Dione as a centralized approach and Delfos as a decentralized approach. We executed experiments to assess the performance of Dione and Delfos.