Algorithms and Systems for IoT and Edge Computing

The idea of distributing the signal processing along the path that starts with the acquisition and ends with the final application has given light to the Internet of Things and Edge Computing, which have demonstrated several advantages in terms of scalability, costs, and reliability. In this dissertation, we focus on designing and implementing algorithms and systems that allow performing a complex task on devices with limited resources. Firstly, we assess the trade-off between compression and anomaly detection from both a theoretical and a practical point of view. Information theory provides the rate-distortion analysis that is extended to consider how information content is processed for detection purposes. Considering an actual Structural Health Monitoring application, two corner cases are analysed: detection in high distortion based on a feature extraction method and detection with low distortion based on Principal Component Analysis. Secondly, we focus on streaming methods for Subspace Analysis. In this context, we revise and study state-of-the-art methods to target devices with limited computational resources. We also consider a real case of deployment of an algorithm for streaming Principal Component Analysis for signal compression in a Structural Health Monitoring application, discussing the trade-off between the possible implementation strategies. Finally, we focus on an alternative compression framework suited for low-end devices that is Compressed Sensing. We propose a different decoding approach that splits the recovery problem into two stages and effectively adopts a deep neural network and basic linear algebra to reconstruct biomedical signals. This novel approach outperforms the state-of-the-art in terms of quality of reconstruction and requires lower computational resources.