Gearbox Anomaly Detection In Wind Turbines Using Classical Machine Learning

Abstract

The primary goal of this master thesis is to develop a data-driven model that can classify or predict the anomalies of wind turbines’ gearboxes by applying machine learning techniques. Firstly, the thesis reviews the development of wind turbines from the origin to the state-of-the-art. Owing to the tremendous demand for wind energy, the turbines move offshore and continue growing in size and number. Consequently, this would result in an increment of operational and maintenance costs, which simultaneously poses many challenges for the wind industry. From the literature review, studies highlighted the drivetrain and gearbox as being responsible for major failures and the downtime of wind turbines. In particular, bearings and gears are the root cause of the recorded damage. This thesis examines the most typical malfunctioning modes of the wind turbine gearbox. The importance and evolution of operational condition monitoring are covered. Subsequently, a data-driven approach is considered, and the possibilities of various experimentations are discussed. A machine learning pipeline with classification algorithms is developed to distinguish between healthy and damaged scenarios. Hyperparameter optimization is introduced and performed to improve the model’s efficiency. The accuracy of the methods applied in this thesis is verified with the actual healthy and damaged vibratory signals from a reference wind turbine gearbox. This can open the gate to further investigations and real-time applications for condition monitoring of wind turbines’ gearboxes under harsh operational environments.