Detecção de desbalanceamento de massa no rotor de turbinas eólicas utilizando algoritmos de aprendizado profundo

Abstract

According to March 18, 2022 data from the Brazilian Wind Energy Association (ABEEólica) the average annual wind power generation supplies about 29 million households per month in Brazil, which corresponds to 86.4 million inhabitants. The country has 795 wind farms, 9,176 wind turbines in operation in 12 states. Wind turbines are exposed to adverse and highly variable weather conditions. Due to these external variations, wind turbines experience constantly changing loads, resulting in operating conditions that lead to intense mechanical stresses. The analysis of the costs related to the operation and maintenance of the existing parks is very important for the parks to continue functioning effectively in the Brazilian energy matrix. Thus, predictive maintenance is one of the best alternatives for the maintenance team to be able to schedule the intervention, avoiding prolonged stops on the production line. Characterized by the measurement and analysis of machine conditions, it is possible to predict possible failures in wind turbines such as the unbalanced mass between the blades, which is one of the most common failures in wind turbines. Thus, the general objective of the work is to develop predictive analysis of mass imbalance failure in wind turbine rotors by applying the Fully Convolutional Network (FCN) and Residual Network (Resnet) in a supervised learning context from the classification of univariate time series of estimated wind turbine rotation speed in the frequency domain generated through a framework using TurbSim/FAST/Simulink of a 1.5 MW wind turbine for different wind scenarios and blade unbalance parameters. Therefore, there was a significant difference in the performance results of the FCN algorithm when using class balancing on the datasets compared to the Resnet algorithm. Furthermore, it can be seen that the FCN algorithm obtained per-class and final performance results with high variability, while Resnet was able to obtain more consistency. Thus, it was possible to conclude that the experiments using balanced class datasets with the Resnet deep learning algorithm was able to obtain the best average performances for the tests performed.