Avaliação das condições de operação das pás de uma turbina eólica em escala: aplicação da DFA combinada com técnicas de reconhecimento de padrões dos sinais de vibração

• Ano de defesa: 2017
• Autor(a) principal: Melo Júnior, Francisco Erivan de Abreu
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Engenharia mecânica; Redes neurais; Vibração; Neural network; Gaussian discriminator; Vibration analysis
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/28386

Wind turbines are machines that operate exposed to adverse climatic conditions of each region and several factors contribute to the blades deterioration. The performance of the wind turbine is directly connected to the aerodynamic efficiency of blades, which are critical components of the project. The turbine’s rotor is a component that suffers constantly with the dynamic loads of the wind. The vibration in a wind turbine is mathematically complex and difficult to be modeled and studies show that conventional vibration analysis techniques are not capable of diagnosing faults present in wind turbines due to the nature of the vibration signals, which are random and non-stationary. This work proposes a study of the vibration signals extracted from a scaled wind turbine, profile NREL S809, diameter of 40 cm and blade tip speed ratio, λ, equal to 7. To simulate potential problems suffered by the blades, masses were added weighting 0.5, 1.0 and 1.5 g to the tip of one blade and later of two blades, to generate from simple imbalances to severe vibration levels in the rotor; beyond the normal condition where the three blades and the system will be balanced. The signals were pre-processed by fluctuation analysis DFA (Detrended Fluctuation Analysis), which in recent years has been widely used to detect properties and long-term correlations in non-stationary time series; then the signals were classified by KLT (Karhunen-Loève Transform), Gaussian discriminator and Artificial Neural Network, all pattern recognition techniques with supervised learning. In general, the classifiers achieved good results in the recognition of the patterns under study, being able to differentiate, with indices greater than 95%, the normal conditions of operation from conditions with imbalances present. In most part of the results, they were also able to classify the different levels of imbalances present in the blades, and may be an excellent predictive maintenance tool in monitoring vibrations of wind turbines.