KOINOS – Knowledge from observations and inference in networks of sensors

This paper fits into the research domain aimed at inferring useful information by mining from data streams of poor quality in wireless sensor networks, with the objective to reduce at the same time the communication load and energy consumption. A set of nodes collect sensor readings from the environment and maintain local model of their evolution. The evaluation of these models is performed using them to simulate data evolution and evaluating the error w.r.t. new sensor readings: when this error for a model exceeds a threshold, model parameters are updated and sent to the sink. At the sink the values collected by sensors are known by using the parameters of local models to simulate sensor’s readings, minimizing the communication among sensors and sink and hence energy consumption. At the sink a global model, a Bayesian Network built on forecasted data, captures spatial and data dependencies among sensors, to detect single sensor and network wide anomalies missed by local error control.