Wind turbine fault detection: a semi-supervised learning approach with two different dimensionality reduction techniques

The quest to save the environment has led many countries to change their mix of energy sources, with most countries focusing on wind energy. Since wind turbines are at the centre of this revolution, ensuring their continuous operation free of faults is a key to success. Towards that end, making use live operational signals from the turbine, various data-driven methods for fault prediction using traditional machine learning have been proposed. This work adopts a novel, automatic, end-to-end AutoML approach covering aspects from features and hyperparameters selection to fault prediction using semi-supervised support vector machines guided by the multi-objective optimisation framework non-dominated sorting genetic algorithm II (NSGA-II). Experiments were carried out using a dataset containing unlabelled records of five 2.0 MW wind turbines. We found that our AutoML approach using NSGA-II for feature selection offers up to 9% improvement in solution quality over the state-of-the-art while being fully automated and requiring no costly and time-consuming feature engineering.