Smart Sensing: Selection, Prediction and Monitoring

A sensor is a device which is used to detect physical parameters of interest like temperature, pressure, or strain, performing the so called sensing process. This kind of device has been widely adopted in different fields such as aeronautics, automotive, security, logistics, health-care and more. The essential difference between a smart sensor and a standard sensor is its intelligence capability: smart sensors are able to capture and elaborate data from the environment while communicating and interacting with other systems in order to make predictions and find intelligent solutions based on the application needs. The first part of this thesis is focused on the problem of sensor selection in the context of virtual sensing of temperature in indoor environments, a topic of paramount importance which allows to increase the accuracy of the predictive models employed in the following phases by providing more informative data. In particular, virtual sensing refers to the process of estimating or predicting physical parameters without relying on physical sensors, using computational algorithms and predictive models to gather and analyze data for accurate predictions. We analyze the literature, propose and evaluate methodologies and solutions for sensor selection and placement based on machine learning techniques, including evolutionary algorithms. Thereafter, once determined which physical sensors to wield, the focus shifts to the actual methodology for virtual sensing strategies for the prediction of temperatures allowing to uniformly monitor uncovered or unreachable locations, reducing the sensors deployment costs and providing, at the same time, a fallback solution in case of sensor failures. For this purpose, we conduct a comprehensive assessment of different virtual sensing strategies including novel solutions proposed based on recurrent neural networks and graph neural networks able to effectively exploit spatio-temporal features. The methodologies considered so far are able to accurately complete the information coming from real physical sensors, allowing us to effectively carry out monitoring tasks such as anomaly or event detection. Therefore, the final part of this work looks at sensors from another, more formal, point of view. Specifically, it is devoted to the study and design of a framework aimed at pairing monitoring and machine learning techniques in order to detect, in a preemptive manner, critical behaviours of a system that could lead to a failure. This is done extracting interpretable properties, expressed in a given temporal logic formalism, from sensor data. The proposed framework is evaluated through an experimental assessment performed on benchmark datasets, and then compared to previous approaches from the literature.