Fault Detection in a Wind Turbine Hydraulic Pitch System Using Deep Autoencoder Extracted Features

A wind turbine is equipped with lots of sensors whose measurements are recorded by the supervisory control and data acquisition (SCADA) system and stored every 10 minutes. The pitch subsystem of a wind turbine is of critical importance as it presents the highest failure rate. Thus, selecting the most essential features from the SCADA system is performed in order to detect faults efficiently. In this study, a feature space of 49 features is available, referring to the condition of a hydraulic pitch system. The dimensionality of this feature space (original input space) is reduced using a Deep Autoencoder in order to extract latent information. The architecture of the Autoencoder is investigated regarding its efficiency on fault detection task. This way, effect of new extracted features on the performance of the classifier is presented. A Support Vector Machine (SVM) classifier is trained using a set of healthy (fault free) and faulty data, representing different kind of pitch system failures. The data are acquired from a wind farm of five 2.3MW fixed-speed wind turbines. The performance metric used to evaluate their effect on data is F1-score.  Results show that SVM using new extracted feature by Autoencoder outperforms SVM classifier using the original feature set, underlining the power of Autoencoders to unveil latent information.