Creative Commons

This dissertation was written as a part of the MSc in ICT Systems at the International Hellenic University. It investigates the utilization of artificial intelligence and specifi-cally artificial neural networks, aiming to increase the robustness of systems that depend on inconstant IoT sensor data for critical operation. The objective is the generation of reliable data stream after a sensor failure or malfunction for a sufficient timeframe that would enable the system to operate without obstruction until all necessary repairs or re-placements occur. A good candidate system is that of an IoT enabled meteorological station. The measurements of various meteorological parameters depend on sensors that due to exposure to the elements are more susceptible to failures. In the case of a sensor failure, the user gets notified and run a previously trained neural network model in order to forecast the values of the failed sensor from the combination of other sensor data un-til the failed sensor could be replaced or repaired. Although the operation of the station could continue with the data of the failed sensor missing, this is a good test case to demonstrate the ability of neural networks to substitute faulty sensors reliably and could be implemented to more sensor-reliant systems as well. The implementation of the projects is divided into two parts. The first part is the build-ing of an autonomous IoT meteorological station complete with user interface and ex-ternal data storage. Besides the time and budget restriction, an effort has been made in order for the station to meet certain standards, so the collected data are reliable and con-sistent. Therefore, a cross-check could be made with other stations in the vicinity. The second part is the data prediction part based on the data from the station using sev-eral tools. Data prediction is a multi-step process that requires data acquisition and preprocessing like cleansing and normalization, and thereafter the use of these data to implement and train an artificial neural network model. The goal is to achieve substitute values with minimal deviation from the real values that could enable reliable operation of the system.