Diagnóstico de conversores CA-CC-CA através da aplicação de redes neurais convolucionais

Abstract

Due to the need to reduce emissions of polluting gases and the search for more sustainable means of energy production, wind energy has been one of the fastest-growing renewable sources in the world. However, operation and maintenance costs can reach up to 15% of the generated energy price, highlighting the need for the application of methods for correct fault diagnosis, especially in the conversion system, which is responsible for about 13% of turbine failures per year. Therefore, this work aims to diagnose the faults presented by power converters in a wind generation plant. For this purpose, a Convolutional Neural Network (CNN) ResNet50 was implemented with real input data from the plant represented as radar graphs and divided into 48 classes: one for the system in normal operating conditions, another for other types of faults, and the rest each representing a fault in the conversion system. As for the results, they outline the progress and challenges faced in the development of the neural network, achieving a final accuracy of 91.66% through the adjustment of hyperparameters and analysis of the complexity of the patterns learned by the network. Consequently, from the observation of the results, it was seen that one of the main offenders for the increase in the model’s accuracy is data imbalance, suggesting the use of techniques such as data augmentation to generate new synthetic examples aiming to balance the distribution. Furthermore, this work not only validates the effectiveness of the proposed model but also sheds light on promising future directions, expanding the debate on the use of neural networks for fault diagnosis in converters, in order to contribute to the increased availability of the generation plant