Posters

Abstract

The integration of Artificial Intelligence (AI) and Machine Learning (ML) in managing and maintaining wind turbine fleets has shown promise for improving operational efficiency and performance. However, the success has been limited and varies depending on the specific use case. One area where AI and ML has underdelivered is within Condition Monitoring Systems (CMS). Typically, CMS record and analyse vibrations in the drivetrain components, most commonly gearbox and main bearing. The amplitude and frequency of these vibrations are monitored over time, with unusual or increasing trends indicating potential damage, misalignments and other mechanical issues. This allows maintenance to be planned and completed in advance of major component issues, reducing the maintenance cost for operators. With current state-of-the-art tools, analysis of CMS data requires significant time from a highly trained engineer, due to the technical nature of the analysis and significant cost associated with missed detections. As the number of installed wind turbines accelerates, this poses potential risks for the industry, as there will be an increasing demand for skilled engineers to maintain the current standard of fault detection. However, innovative solutions can be applied to allow monitoring engineers to work at a larger scale while maintaining quality. To address these challenges, this paper presents three case studies for innovative statistical and ML solutions to further automate the analysis of CMS data. These case studies, grounded in real-world data and application, demonstrate the practical effectiveness of integrating statistical methods and machine learning algorithms in enhancing the predictive accuracy and operational efficiency of Condition Monitoring Systems for wind turbines. The first study focuses on reducing false positives by identifying and classifying bad data or missing data. The second explores pattern detection in vibration spectra as an enhanced decision support tool. The third proposes an automated fault diagnosis method. The developed approaches significantly improve user efficiency, reducing the engineer time spent managing alarms as well as the timeliness of fault detection while maintaining high accuracy metrics, thereby reducing O&M costs and ultimately the Levelised Cost of Energy (LCOE). Furthermore, the challenges in implementing these technologies are addressed, including data collection, model training, model generalisation and integration with existing turbine monitoring systems. Recommendations are made to address these challenges. In conclusion, this paper underscores the transformative potential of integrating numerous statistical and ML techniques in CMS data analysis, demonstrating their compounding effect. These advancements not only optimise maintenance practices but also play a crucial role in the sustainable expansion of wind energy.