Posters

Abstract

As wind energy assets age, the implementation of advanced predictive maintenance strategies is essential to mitigate unplanned downtimes and optimize Operation and Maintenance (O&M) costs. The pitch system is a critical subassembly, frequently identified as the component with the highest failure rate, often accounting for over 20% of total turbine downtime in some studies (Carroll et al., 2016). This study explores a successful application of machine learning (ML) models—based on Artificial Neural Networks (ANN)—to monitor the end-of-life phase of pitch bearings in a 1.6 MW wind turbine with an electrical pitch system and  by detecting anomalies in pitch motor electrical currents. The predictive model established a baseline for healthy pitch motor currents, facilitating the early detection of increased mechanical resistance stemming from bearing degradation. In this study, the model identified anomalous behavior in Blade 1 on March 4, 2024, followed by Blade 2 on October 24, 2024. The subsequent bearing replacement was executed in April 2025, demonstrating a lead time of 399 days for Blade 1. This significant window proved critical for optimizing supply chain logistics and proactive maintenance scheduling. The impact of the degradation period was severe. During the alert phase, the turbine accumulated 568.97 hours of pitch-related downtime, representing 62.9% of the total downtime for the asset. This resulted in a loss of 618.91 MWh, which accounted for 66% of the energy lost due to downtime and 8.74% of the turbine's total gross production in that period. Immediately following the intervention, the machine learning models originally employed for anomaly detection were repurposed as a validation framework to confirm the restoration of healthy motor current behavior. This approach proved instrumental, as direct comparisons of raw data are often confounded by varying operational conditions. By normalizing these environmental and operational variables, the predictive model enables a definitive verification of component health within a 24-hour observation window, ensuring the maintenance successfully restored the system to a healthy baseline. Post-repair analysis confirmed the model's effectiveness as a validation tool. Following the intervention, predictive alarms ceased. A comparative analysis of the 9-month period post-repair showed a significant reduction in the impact of pitch-related failures. The ratio of pitch-related energy loss to total production dropped from 8.74% to 4.53%. This represents a net gain of 4.21% in total turbine production, during the post-repair period, due to the reduction in frequency and severity of pitch system downtime events. The findings substantiate that machine-learning-driven monitoring of electrical current anomalies serves as a high-fidelity proxy for mechanical health. This methodology facilitates the strategic optimization of pitch bearing end-of-life management and provides a robust framework for post-maintenance validation.  References: * Carroll, J., McDonald, A., & McMillan, D. (2016). "Failure rate, repair time and unscheduled O&M cost analysis of offshore wind turbines." IEEE Transactions on Sustainable Energy. * Zhang, Z. (2018). "Automatic Fault Prediction of Wind Turbine Main Bearing Based on SCADA Data and Artificial Neural Network." Open Journal of Applied Sciences.