Hydrogenerators are crucial assets for both power generating companies and the population that relies on this supply. However, these machines are susceptible to various types of defects that can result in unexpected interruptions if nothing is done about it. Partial discharge analysis is an already established approach for assessing the condition of high-voltage equipment, with the automatic detection of different types of defects being essential, since different levels of risk to operation vary according to the type of discharge. Deep neural networks have been proposed aiming at the classification of partial discharges using phase-resolved partial discharge (PRPD) diagrams. However, obtaining labeled datasets with a large number of examples is a problem that directly impacts the performance of supervised trained classifiers. In this context, this dissertation proposes a semi-automatic technique for PRPD labeling, based on dimensionality reduction strategies and data clustering, as well as investigates the use of Generative Adversarial Networks (GANs) in artificially expanding the training set. The dataset used in the work consists of real PRPDs obtained through online monitoring procedures of partial discharges in hydrogenerators. The performance of the proposed systems is contrasted with recent representative results from the state-of-the-art in the field. The results demonstrate that the application of the proposed semi-automatic labeling technique has the potential to considerably reduce the workload and time associated with manual classification. Furthermore, the use of artificial PRPDs generated by GANs resulted in a notable improvement in the classifier’s performance, reaching an average accuracy of 94.72%, compared to 89.44% obtained with the best competing technique. Similar gains were also observed in class accuracies.
|
