A Hybrid Method for Fault Detection of Wind Turbine Drivetrains Using Machine Learning and Signal processing.

Description

In recent years, the increasing public interest in renewable energy sources has directed a growing amount of research to investigate the improvement of its cost efficiency. Wind energy is expected to provide a substantial contribution to renewable energy. However, the operating and maintenance (O&M) cost is a significant part of the total price with the turbine drivetrain being a substantial contributor to this O&M cost. A common method of monitoring the drivetrain is through vibration analysis. These vibrations are analysed through signal processing techniques to enable the detection of mechanical faults trends. However, the current state-of-the-art vibration signal processing techniques produce an excessive number of condition indicators. When the volume of data gets higher on a fleet level, manual indicator monitoring becomes unfeasible. A hybrid fault detection method to leverage state-of-the-art signal processing techniques using machine learning methods can provide high-level health indicators. The signal processing indicators are fused to derive a high-level health indicator. First, an anomaly detection model is trained on healthy historical data to learn each indicator's normal behaviour. This model can identify the indicator's anomalous trends by computing the deviation from the expected behaviour. The high-level health indicator is derived by fusing anomalies over multiple indicators. The end user can get a clear overview of the machine's health status by only observing the high-level health indicator. The proposed method is validated on a complete offshore wind farm fleet having multiple years of condition monitoring data with labelled fault cases. The condition monitoring data includes labelled fault cases in the planetary gear stage, generator failure and a high-speed stage failure. The proposed method enables the end user to get an overview of the machines' health status over the fleet without looking into individual indicator trends. Future work will investigate a multivariate deep learning anomaly detection model to predict a high-level health indicator.

Periode	25 mei 2023
Evenementstitel	Wind Energy Science Conference 2023
Evenementstype	Conference
Locatie	Glasgow, United KingdomToon op kaart
Mate van erkenning	International