Utilização de modelos computacionais de aprendizagem de máquinas para previsão de valores de temperatura do rolamento da multiplicadora de um aerogerador em operação comercial.

Abstract

With the popularization of predictive maintenance, the search for non-destructive testing and Noninvasive inspection techniques have steadily increased in various sectors industry and with this the use of innovative technologies has stood out. The use of ultrasound, thermographic images, vibration analysis and videoscopy are examples of these techniques. Another theme in evidence is the use of computational methods of machine learning. These methods were developed in the 1970s, but due to data processing and storage limitations exploited around the 2000s and has been increasingly used. THE versatility of the methods has led to their implementation in various areas of I study biotechnology, engineering, statistics and marketing. In engineering these methods have been successfully applied for the most diverse objectives and the cost The implementation of these models has been increasingly reduced. The purpose of present work was to program the computational methods of learning of Machines, Linear Regression, Polynomial Regression and Support Vector Machine, developed in the R programming language to a set of data collected from the SCADA system of a wind turbine in commercial operation in order to predict temperature values ​​for the mainshaft bearing of the multiplier, cross the results obtained with the different models and present the comparison between methods. It was concluded that the method with higher assertiveness, 10.3% more than the Linear Regression and 8.96% Polynomial Regression, was the Support Vector Machine with an average error of approximately 1.25ºC for wind turbine 1 and 1.36ºC for wind turbine 2 and a correlation coefficient of 85.6% percent for wind turbine 1 and 88.3% for wind turbine 2.