Detecção de falhas em aerogeradores com análise de temperaturas e modelos estatísticos

Abstract

In a world where the energy demand is constantly increasing, wind power is essential because of its sustainable nature, competitive costs, and outstanding potential to produce energy. However, wind turbines are giant machines with high installation, operation, and maintenance costs. A wrong maintenance strategy could ruin the cash flow of a wind farm, compromising the investment of decades. Wind turbines have monitoring systems for evaluating whether they work within operational limits. These systems store a massive amount of valuable data, which could provide an early indication of faults occurring in the machines and, consequently, avoid unexpected maintenance costs. Temperature is commonly used to monitor the condition of mechanical machinery. The wind turbine supervisory system monitors critical components’ temperature and stores this information in a database. In this sense, the present work proposes developing a machine-learning model based on the analysis of the temperature of wind turbine components to determine, in advance, whether it is operating in its normal behavior. The present methodology proposes implementing normal-behavior models using different machine learning algorithms (artificial neural network, random forest, and K-nearest neighbor) to detect faults in wind turbine components such as the gearbox, main bearing, and generator. Operational data from wind turbines installed in Brazil are used, and challenges about using unlabeled real data are discussed throughout the paper. Some challenges are filtering data, selecting variables and data windows for training, and validating the models. The models can identify deviation from normal behavior, characterizing a fault, even before the supervisory system triggers the alarms. Results also present what methods perform efficiently and if there are differences regarding the analyzed component. Anticipated actions from the maintenance staff to correct the faults can be carried out in a planned and efficient way, which not only preserves the wind turbine but also increases the wind farm’s key performance indicators. Different detection periods were identified depending on each component’s dynamics and the model’s particularities. The better performing models were artificial neural networks and decision trees, detecting faults from 80 to 100 days in advance for the gearbox and 90 to 120 days before for the main bearing.