GRAPH SIGNAL PROCESSING WITH AN APPLICATION TO OVERHEATING DETECTION IN ELECTRICAL ENERGY SWITCHBOARDS .
Signal processing. Graphs. Sensor networks. Anomaly detection. Electrical installations.
In recent decades, several technological advances have allowed the massive production and storage of data related to a variety of real-world scenarios. Many of these scenarios can be represented as networks over which the respective data is distributed. This is the case, for example, of a network of spatially arranged sensors dedicated to peforming measurements that are interrelated in some sense, composing what would be a signal over the network in question. In general, scenarios like the one exemplified can be modeled as a graph whose nodes are associated with samples of a signal. To deal with signals of this nature, graph signal processing (GSP) was proposed, which seeks to extend to the so-called vertex domain concepts and operations of classical signal processing, dedicated to analyzing signals in domains such as discrete-time only. In this dissertation, a review on the fundamentals of graph signal processing is presented, with emphasis on applications of the respective theory in problems related to sensor networks. As an original contribution of this work, a case study is carried out, which consists of applying GSP to detect overheating in electrical energy switchboards. This study considers the electrical network of a large hospital in the metropolitan region of Recife, modeling it as a graph whose nodes correspond to the energy switchboards. Despite limitations mainly related to the restricted amount of data available for the study, the results obtained suggest that the GSP can be a useful tool for the application in question, providing satisfactory evidence about the appearance of hot spots in the analyzed network.