FAULT DETECTION AND DATA PREDICTION FOR WIRELESS SENSOR NETWORKS

In the last few years, Wireless Sensor Networks (WSNs) have been extensively used as a pervasive sensing module of Ambient Intelligence (AmI) systems in several application fields, thanks to their versatility and ability to monitor diverse environmental quantities. Although wireless sensor nodes are able to perform onboard computations and to share the sensed data, they are limited by the scarcity of energy resources which heavily influences the network lifetime; moreover, the design phase of a WSN requires testing the application scalability prior to actual deployment. In this regard, this dissertation focuses on data prediction to address such crucial tasks as prolonging the network lifetime and testing the WSN scalability. Nevertheless, the matter is particularly challenging as the real world measurements are influenced by unpredictable events that affect the sensor readings. To this aim, fault detection techniques help to identify corrupt measurements and to discard them before they are actually transmitted within the network, so they may be profitably used to improve the precision of the prediction models. This dissertation describes the design of two software modules which address fault detection and data prediction and may be combined in a single software system for WSNs. The fault detection submodule classifies the sensed measurements as “corrupt” or “regular” by means of Bayesian Inference. The prediction submodule builds models for the monitored quantities and is also able to generalize them to unknown environments populated by virtual sensor nodes so it allows to test the scalability of the application for networks of different sizes. Prediction also allows sensor nodes to reduce their energy consumption as much as possible by fine tuning their sampling rate based on the accuracy of the predictors. Experimental results show the capabilities of the proposed system to detect faults and to build reliable prediction models for some of the most common physical quantities for WSNs, namely light exposure, temperature and humidity.