Leveraging signal processing and machine learning for automated fault detection in wind turbine drivetrains

Abstract. Wind energy is considered a sustainable renewable energy source; however, it faces the challenge of significant operating and maintenance costs. The research proposes a hybrid fault detection method to combine the physical domain knowledge with the machine learning models to provide an overview of the health of wind turbine drivetrain components. Signal processing indicators are computed from raw vibration signals measured from strategically placed accelerometers over drivetrain components. It produces an immense number of indicators as each indicator is sensitive towards certain types of faults, and manual monitoring becomes an unfeasible task. The machine learning models are trained using signal processing indicators and supervisory control and data acquisition (SCADA) data. The normal behavior modeling technique is employed to learn the healthy operation of the machine from data collected during healthy machine operation. The trained normal behavior machine learning models label each indicator in a healthy or faulty state over time. The labeled state-of-the-art signal processing indicators are fused to provide a high-level health status overview of wind turbine drivetrain components. It helps to derive the required details from many condition indicators, which is valuable when managing multiple components in a single wind turbine across an entire wind farm. The proposed hybrid fault detection method is validated on an offshore wind farm with multiple years of condition monitoring data. It provides a high-level health overview that is readily understandable for non-expert wind farm operators, and for more detailed fault analysis, experts can conduct a comprehensive inspection.