Ensemble learning framework for fleet-based anomaly detection using wind turbine drivetrain components vibration data.

Anomalies in wind turbines pose significant risks of costly downtime and maintenance, underscoring the importance of early detection for reliable operation. However, conventional fault detection methods, often reliant on standalone anomaly detection models, struggle with generalization in such complex settings, leading to suboptimal prediction performance. To address this challenge, this proposes an ensemble technique pipeline to enhance robustness by combining multiple models for anomaly detection using condition monitoring system vibration data from selected wind turbine bearings. A fleet-based anomaly detection framework was applied and improved into a comprehensive ensemble pipeline. Thus, the novelty of this study lies in the in-depth evaluation of using ensemble techniques with anomaly detection models for condition monitoring system vibration data, providing insights into the effectiveness of such an approach. In the end, the proposed pipeline attained over 84% for the receiver operating characteristic curve (AUC) across components when deployed over real unseen data, achieving 98% for AUC for the main bearing through Majority-Ensemble, 89% for AUC for the gearbox high-speed shaft bearing under key nearest neighbor stacking, 84% for AUC for the generator drive-end bearing through Voting-Hard technique and 95% for AUC for the generator non-drive-end bearing under Voting-Soft. This study demonstrates ensembles can achieve robust anomaly detection for wind turbine components, addressing generalization challenges when backed by robust pipelines.