Search

Recently, several researches have been carried forward about self-supervised learning, with the goal of remedying the necessity of large amount of labeled data that characterises supervised deep learning approaches. Indeed, although supervised learning methods may achieve impressive performance, labeling huge amount of data is often impracticable since it may require expert knowledge and/or data may be scarcely available. Through the use of an auxiliary task, called pretext task, a self-supervised model can be trained without the need of manual annotations using as pseudo-labels information obtained directly from the data. Typically, the pretext task is defined to give a capability to the network useful for the final task, also called downstream task, to which it should be correlated. This project aims to develop a multimodal self-supervised method, that learns an embedded space for audio and visual video data to provide a reasonable feature representations of specific actions in the EPIC Kitchen Dataset. Using the concatenation of two different proxy tasks, mainly audio-video synchronization and audio-video alignment, a deep learning model can obtain strong temporal reasoning capacity able to describe an action in a multimodal space used help to represent the main features of the action. To validate the quality of the learnt representation, a fine-grained cross-retrieval task for audio and videos is performed on the features extracted from the entire dataset. Moreover some classifiers are trained on the same features to test the capacity of the embedded space in term of action recognition. This thesis was developed in collaboration with IRII - Institut de Robòtica i Informàtica Industrial in Barcelona, Spain, as final project of the master in Automation Engineering, held at University of Padova, Italy

Wireless power transfer networks (WPTNs) are composed of dedicated energy transmitters (ETs) that charge energy receivers (ERs) via radio frequency waves. A safe-charging WPTN should keep electromagnetic radiation below predetermined limits meanwhile maximizing the transmitted power. In this paper, we consider this requirement as an optimization problem: the maximization of harvested power by ERs subject to the electro-magnetic safety constraints. In order to provide an approximated solution to this problem, we introduce a dual ascent-like distributed charging algorithm that enables ETs to work without global information and satisfy safety constraints asymptotically. We provide an in-depth theoretical analysis of our algorithm which is supported by numerical simulations., Accepted author manuscript, Embedded and Networked Systems

This paper proposes a technique to control distributed energy resources in low-voltage microgrids aiming at (i) allowing power flow control at the point of connection with the upstream grid, (ii) keeping voltage profiles within the operational limits. The first feature is crucial in smart low-voltage power systems. In fact, it enables both demand-responses, which is extremely valuable from the point of view of distribution system operators and for energy trading, and the autonomous operation of the microgrid. The latter can be achieved by regulating to zero the power exchanged with the main grid. The second feature allows to limit voltage increases due to active power injection by distributed energy resources and, thus, to limit stresses on the electrical infrastructure and the served loads, which is a concrete issue as renewables become widely deployed in the low-voltage scenario. The proposed approach is firstly described in detail, then a systematic analysis of its local and global properties is reported. All the obtained results are verified considering the IEEE 37 test feeder in realistic operating conditions.

This thesis proposes distributed algorithms to automatically deploy a group of robotic agents and provide coverage of a discretized environment represented by a graph. The algorithms presented converge to a centroidal Voronoi partition using only pairwise gossip communication between the agents. The communication protocols for the algorithms are Synchronous Gossiping communication and Asynchronous Broadcast communication. Both algorithms have been implemented by the author of the thesis in MATLAB. This thesis proves that both algorithms make the territory ownership to converge to a pairwise-optimal partition in finite time. Additionally, some simulations are done in MATLAB proving how the Asynchronous Broadcast coverage algorithm reduces the amount of communications for large teams in large environments.