In the context of fault detection of bearings, state-of-the-art unsupervised methods struggle to capture anomalies when both speed and loads are a variable. This limitation may arise due to the lack of knowledge about a possibly faulty system behavior and the reliance on a fixed operating point of the baseline system. In this work a machine learning (ML) approach is implemented to address these challenges. To this end, an autoencoder framework is used to extract fault-sensitive features, a support vector machine is employed to classify them, and a pseudo-anomaly scheme is used to ensure a balanced training dataset with an equal distribution of healthy and faulty data. The application concerns a laboratory experiment of a progressively damaged bearing in a wind turbine drivetrain simulator. Different operating points of the drivetrain system related to changing speed and load are considered. The result highlights the potential of ML-based methods for fault diagnosis of wind turbine bearings, offering a more reliable alternative to the conventional fault detection methods.