Citation: Korkos, P.; Kleemola, J.; Linjama, M.; Lehtovaara, A. Representation Learning for Detecting the Faults in a Wind Turbine Hydraulic Pitch System Using Deep Learning. Energies 2022, 15, 9279. https://doi.org/10.3390/ en15249279 Academic Editor: Davide Astolfi Received: 8 November 2022 Accepted: 4 December 2022 Published: 7 December 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affil- iations. Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). energies Article Representation Learning for Detecting the Faults in a Wind Turbine Hydraulic Pitch System Using Deep Learning Panagiotis Korkos 1, * , Jaakko Kleemola 2 , Matti Linjama 3 and Arto Lehtovaara 1,† 1 Tribology and Machine Elements, Materials Science and Environmental Engineering, Faculty of Engineering and Natural Sciences, Tampere University, P.O. Box 589, 33014 Tampere, Finland 2 Suomen Hyötytuuli Oy, P.O. Box 305, 28601 Pori,Finland 3 Automation Technology and Mechanical Engineering Unit, Faculty of Engineering and Natural Sciences, Tampere University, P.O. Box 589, 33014 Tampere, Finland * Correspondence: panagiotis.korkos@tuni.fi † Retired. Abstract: Wind turbine operators usually use data from a Supervisory Control and Data Acquisition system to monitor their conditions, but it is challenging to make decisions about maintenance based on hundreds of different parameters. Information is often hidden within measurements that operators are unaware of. Therefore, different feature extraction techniques are recommended. The pitch system is of particular importance, and operators are highly motivated to search for effective monitoring solutions. This study investigated different dimensionality reduction techniques for monitoring a hydraulic pitch system in wind turbines. These techniques include principal component analysis (PCA), kernel PCA and a deep autoencoder. Their effectiveness was evaluated based on the performance of a support vector machine classifier whose input space is the new extracted feature set. The developed methodology has been applied to data from a wind farm consisting of five 2.3 MW fixed-speed onshore wind turbines. The available dataset is composed of nine pitch events representing normal and faulty classes. The results indicate that the features extracted by the deep autoencoder are more informative than those extracted by PCA and kernel PCA. These features led to the achievement of a 95.5% F1-score, proving its superiority over the traditional usage of original features. Keywords: pitch system; wind turbine; SCADA; fault detection; feature extraction; deep autoencoder 1. Introduction Nowadays, wind farms are critical infrastructures for every country, especially with the power production market changing because of Russian gas restrictions. The power produced by wind farms is expected to balance the demand and keep the levelised cost of energy as low as possible. However, wind farm production depends on the wind conditions and the availability of wind turbines, meaning that they should be fault-free for when there are ideal conditions for power production. The latter is the only factor that can be controlled by humans; thus, the condition monitoring of wind turbines is crucial to ensure their availability. As a result, predictive maintenance can also lower the cost of the produced energy. According to the latest data from Danish wind farms [1,2], the operation and maintenance (O&M) costs of a wind turbine, regarding the market price, correspond to approximately 1.3–1.6 euro cents/kWh to ensure the profitability of the asset. The levelised cost of energy produced by an onshore wind turbine is assumed to be between 3.94 and 5.01 euro cents/kWh when under favourable wind conditions [3]. For this reason, operators are urged to lower these costs by utilising more sophisticated tools for scheduling their maintenance tasks. The Supervisory Control and Data Acquisition (SCADA) system is a key element in achieving these goals. Every wind turbine, regardless of the manufacturer, is equipped with a SCADA system whose task is to store the sensor measurements in a Energies 2022, 15, 9279. https://doi.org/10.3390/en15249279 https://www.mdpi.com/journal/energies